System for reconstructing the 3-dimensional motions of a human figure from a monocularly-viewed image sequence

Abstract

A system is provided for reconstructing the 3-dimensional motions of a hu figure from a monocularly viewed image sequence in which a statistical approach is used coupled with the use of a set of motion capture examples to build a gaussian probability model for short human motion sequences. In a simplified rendering domain, this yields an optimal 3-d estimate of human motion used to create realistic animations given a 2-d temporal sequence. The subject system is also useful to identify which motion modes are difficult to estimate. In one embodiment, a stick figure is overlaid on the input video image to allow manual correction of incorrectly tracked body parts. The utilization of stick figure correction permits manual correction for a more accurate 3-d reconstruction of the motion depicted in the input video sequence. An interactive tracking system processes real video sequences, and achieves good 3-d reconstructions of the human figure motion.

Description

FIELD OF THE INVENTION

This invention relates to a system for recognizing body motion and more particularly to a system for reconstructing an estimate of the 3-dimensional positions of a human body from a live or pre-recorded sequence of images of the body.

BACKGROUND OF THE INVENTION

As one watches a film or video of a person moving, one can easily estimate the 3-dimensional motions of the moving person from watching the 2-d projected images over time. A dancer could repeat the motions depicted in the film. Yet such 3-d motion is hard for a computer to estimate.

Many applications would follow from a computer with the same abilities to infer 3-d motions. There are applications to public safety for elevators and escalators, as well as in interactive games, and virtual reality. In computer graphics, a growing industry is devoted to "motion capture", where digitized human figure motion drives computer graphic characters. The human's 3-d motion information is digitized either by magnetic sensors or by optical techniques with multiple calibrated cameras and a special suit of markers. Unfortunately, either technique is expensive and cumbersome. To obtain 3-d figure motion information from single-camera video would allow motion capture driven by ordinary monocular video cameras, and could be applied to archival film or video.

As described by L. Goncalves, E. D. Bernardo, E. Ursella, and P. Peronra, Monocular tracking of the human arm in 3d., Proc. 5th Intl. Conf. on Computer Vision, pages 764-770, IEEE, 1995, under constrained viewing and motion conditions, Goncalves and collaborators, tracked the motion of an arm in 3-d. In an article by J. M. Rehg and T. Kanade entitled Model-based tracking of self-occluding articulated objects, Proc. 5th Intl. Conf. on Computer Vision, pages 612-617, IEEE, 1995, some hand motions are tracked over 3-d, allowing significant occlusions. However, this requires 3-d model initialization, and controlled viewing conditions. Work at recovering body pose from more than one camera has met with more success as discussed by D. M. Gavrila and L. S. Davis, 3-d model-based tracking of humans in action: a multi-view approach, in Proc. IEEE CVPR, pages 73-80, 1996. Despite research attention as illustrated in a book edited by I. Essa, entitled International Workshop on Automatic, Face- and Gesture- Recognition, IEEE Computer Society, Killington, Vt., 1997, the problem of recovering 3-d figure motion from single camera video has not been solved satisfactorily.

SUMMARY OF INVENTION

In order to detect human motion without the use of markers on an individual and without using as many as 6 cameras, an optimizer is used which takes a video input signal and an original training set and outputs optimized 3-d coordinates of body markers corresponding to the video input, the training data and manually input corrections.

In order to accomplish the optimization, the optimizer finds the most probable linear combination of training data to explain the input video, accomodate user corrections, and be a likely 3-d motion segment. Having derived optimal 3-d coordinates, in one embodiment, realistic cartoon characters can be generated, all without the use of physical markers on the actor's body, and all from a single camera viewing the actor. Not only can the motion of an actor be characterized, archival film can be analyzed as to the motions of actors therein. The result is the ability to create animation for whatever purpose in a simple non-cumbersome manner.

In one embodiment, the optimizer performs its function in accordance with the following formula: ##EQU1## where E(.alpha.) is the energy function to be minimized by the optimal coefficients .alpha. to be found, R is the vector of sensor responses over time from the image data. The function f converts .alpha. body motion coefficients to predicted sensor responses. I.sub.i is the ith point position specified bye the user, and P.sub.i projects the .alpha. coefficients onto the corresponding ith stick figure part 2-d position. .lambda..sub.1 and .lambda..sub.2 are constants which reflect the weights of the image data, the priors over human motions, and the interactively specified 2-d point matches.

In the subject invention strong prior knowledge about how humans move in the form of training data is used. This prior knowledge dramatically improves the 3d reconstructions, with the subject prior knowledge being fit from examples of 3-d human motion.

In the subject invention the 3-d reconstruction in a simplified image rendering domain utilizes a Bayesian analysis to provide analytic solutions to fundamental questions about estimating figural motion from image data. Applying the Bayesian method to real images premits reconstruction of human figure motions from archival video, with the subject system accomodating interactive correction of automated 2-d tracking errors that allows reconstruction even from difficult film sequences. In one embodiment, a stick figure is overlain on an image of a moving human to allow the user to ascertain and correct 2-d motion estimate errors.

In summary, a system is provided for reconstructing the 3-dimensional motions of a human figure from a monocularly viewed image sequence in which a statistical approach is used coupled with the use of a set of motion capture examples to build a gaussian probability model for short human motion sequences. In a simplified rendering domain, this yields an optimal 3-d estimate of human motion used to create realistic animations given a 2-d temporal sequence. The subject system is also useful to identify which motion modes are difficult to estimate. In one embodiment, a stick figure is overlaid on the input video image to allow manual correction of incorrectly tracked body parts. The utilization of stick figure correction permits manual correction for a more accurate 3-d reconstruction of the motion depicted in the input video sequence. An interactive tracking system processes real video sequences, and achieves good 3-d reconstructions of the human figure motion.

BRIEF DESCRIPTION OF DRAWINGS

These and other features of the Subject Invention will be better understood in conjunction with the Detailed Description taken in conjunction with the Drawings of which:

FIG. 1 is a block diagram of a system having a video or archival film input for providing 3-d motion estimates used for providing realistic animation.

FIGS. 2a, b, and c are diagrammatic representations showing example motion sequences from a training set of 10 sequences;

FIGS. 3a and b are diagrammatic representations showing approximation of the human motion signal as a linear combination of basis functions, with FIG. 3a showing a 40 basis functions approximation and with FIG. 3b showing the original human motion sequence;

FIG. 4 is a graph showing prior and posterior covariance matrix diagonal elements, plotted on a semi-log scale, with measurment of the marker positions in the image plane dramatically reducing the prior uncertainty of the human motion;

FIGS. 5a, b, and c are diagrammatic representations showing three random draws from the gaussian prior distribution over the 37 3-d marker positions, with the results all looking human, and corresponding roughly to human motions;

FIGS. 6a-d are diagrammatic representations showing the original 3-d sequence, the orthographically projected image with markers connected by lines for clarity, 3-d reconstruction omitting prior information, with the representation being the 3-d figure in the eigenspace of the human motions which best accounts for the image data and a full Bayesian reconstruction, noting that the addition of prior information creates a reconstruction more similar to the original, and with the high posterior covariance modes explaining the remaining differences from the original sequence;

FIGS. 7a-c are diagrammatic representations showing the three modes with the highest posterior uncertainty;

FIGS. 8a-c are respectively: Samples from a 100 frame sequence of Barishnikov dancing with tracked 2-d stick figure overlaid, inferred 3-d marker positions using the gaussian prior model for 3-d figure motions, and the recovered 3-d moving cylinder figure which captures the 3-d motions of the dancer.

DETAILED DESCRIPTION

Refering now to FIG. 1, a system 10 for providing a 3-dimensional estimate of body motion of an individual is used in one embodiment for providing cartoon characters whose motions resemble the motions of an actor. As discussed previously, it is only with difficulty that one can generate realistically moving cartoon characters based on human movement.

In the embodiment illustrated, an actor 12 has his movements filmed by a video camera 14 which are digitized at 16 and coupled to an optimizer 20, the purpose of which is to find the optimal 3-d body coordinates, given a 3-d motion training data set 22.

In one embodiment, the training data is divided up into snippets or segments which define so called basis vectors for human motion. Multiplying these values by coefficients and adding the results yields the new motions. The basis vector is a vector having the following components: a frame number, the number of each marker in that frame, and the x, y, and z positions of each marker. Note that the markers are obtained from the training data set. The markers in the training set indicate the positions of various positions on the body which describe the body part position. These are similar to placing physical markers on the clothing of the individual and measuring the 3-d position of those physical markers by triangulating from multiple cameras. Training data for animation is available commercially from Adaptive Optics Associates, Cambridge, Mass., in which datasets are provided with respect to for instance a dancer, indicating body part position in terms of markers. These example motion sequences in the form of a training set provide the a priori data used in subject system.

It will be noted that a snippet has a pre-determined number of frames. Moreover, in one embodiment there is an offset which describes how many frames are skipped between the start of one snippet and the next.

It is the purpose of optimizer 22 to find the set of coefficients of the 3-d motion training data set which (1) best explains the input video data, and (2) accounts for user corrections which will be explained hereinafter. Note that the subject system operates on 2-d data, which is normally insufficient to provide the 3-d data necessary to describe the body motion to be imparted to a cartoon character.

The 3-d estimates are derived from the 2-d video data through finding the most probable linear combination of the 3-d training data that explains the 2-d image information and the 2-d user corrections.

In order to complete the 3-d description of the body motion, it is necessary to guess the 3rd dimension. How this is accomplished in general is to start with the 3rd dimension information from the training data. Specifically, a prior preference for particular linear combinations of the training data is formed. This is obtained from the training data. Then optimizer 20 trades off the stength of the prior preference for a given linear combination against how well that linear combination explains the input video data and accounts for the user corrections.

Optimizer 20 iteratively searches for an optimum linear combination of training motion snippets and outputs the optimum set of coefficients for the motion under study. As illustrated at 24, this output is referred to herein as the estimated 3-d coordinates of the aforementioned body markers.

While not required, how accurate the estimation is is ascertained in one embodiment by the user in which a 3-d to 2-d projection of the recovered marker positions as illustrated at 26 is transformed into a stick figure which is made to overlay the corresponding video as illustrated at 28 where a visual comparison is made at monitor 30. The user compares the position of the stick figure with that of the actor in the video to ascertain what adjustments must be made to the stick figure in order to bring it into line with the corresponding body positions of the actor. This is accomplished by unit 32. A mouse 34 is utilized in one embodiment to designate what body part of the stick figure is selected, and where that stick figure body part should go. This is accomplished through the selection by the mouse of the marker to be adjusted, and designation by a second mouse click as to where that marker should be moved on the video body image.

Because of the operation of optimizer 20, it is not necessary or desireable to manually adjust each marker. Rather, through an iterative process of specifying selected markers for adjustment, optimizer 20 outputs a revised stick figure estimate. Because there are only certain things that a human body can do, there is a limited range of possibilities for the stick figures. This is because a certain amount of a priori knowledge which is implicitly contained in the training data trickles down to the resultant 3-d stick figure coordinates.

Rather than relying on the placement of markers all over the actors body, the subject system recovers the same 3-d information just from the video sequence of the unencumbered actor. No longer must the actor be viewed from 6 cameras and be provided with apparatus on his clothing to be able to detect his motions. Rather, digitized video is analyzed to detect his motions and provide a 3-dimensional stick figure representation including markers which is used in the generation of corresponding cartoon characters mimicking the motion of the actor.

The estimated 3-d coordinates are used in one embodiment to generate a character model by the fitting of a predetermined character model to these 3-d coordinates, as illustrated at 40. Once the character model has been fit to the 3-d coordinates, the character model is displayed at 42. Note that the display of the character may be altered for desired costume, viewpoint of the camera, so as to rotate the figure, or have the figure interact with other characters or scene elements as in a game. The result is that video game generation is simiplified and made more realistic without cumbersome apparatus on the actor and without costly motion capture equipment.

It will be appreciated that not only can the motion of an actor or athelete be captured by a video camera, archived films or videos can be analyzed so that characters resembling the character on the film or video can be made to move in the same manner as the character on the film or video. Thus the character can be re-rendered in a new setting or from a new camera angle. Moreover, characters originally caught on film can be made to interact with each other without the necessity of physically placing markers on them which would be impossible. One could therefore ressurrect deceased actors and have them interact in a realistic way with characters in a virtual world using the subject technique.

The same techniques can be utilized to ascertain in a given environment what motions constitute non-normal behavior. For instance, people descending on an escalator can be viewed in real time and should an individual fall, this fact can be detected by the subject apparatus. Moreover, time and motion studies can be generated with the subject techniques.

As will be described hereinafter, the function that the optimizer 20 optimizes can be interpreted as the Bayesian posterior probability of a given 3-d body motion. How this is accomplished is through the utilization of a prior probability for a set of snippet coefficients. It will be appreciated that one represents motions as linear combinations of snippets. The prior probability is thus a mathematical expresion of the probability of a human making the motion described by a given set of snippet coefficients. It will be further appreciated that the prior probability comes from the training data absent the new video, eg, prior to looking at the video. The posterior probability refers to the probability of the set after examining the video data, and optionally incorporating the user's corrections.

Optimizer 20 thus utilizes Bayesian mathematical probability techniques to make sure that the 3-d esimates which are used to describe body motion are optimal. This gives an optimal tradeoff between explaining the video data, accounting for the user's corrections, and estimating likely human 3-d motions.

More specifically, in one example the training examples used herein are 10 3-d motion capture sequences, of 5-10 seconds each, obtained commercially. The data is position information of 37 markers over 120 to 240 temporal frames for each sequence, sampled at roughly 20 frames per second. The motions are an ecclectic set of short activities, presumably designed to illustrate the range and precision of the motion capture equipment.

FIG. 2 shows subsets of 3 of the 10 motion sequences. As in other motion displays, these figures show a "stroboscopic" display of an animated sequence, temporally subsampled and sometimes spatially offset for clarity. Lines are drawn between markers or circles to create a stick person that is easier to interpret.

It is the purpose of the subject invention to seek a simple and tractable, yet useful, probabilistic model for the 3d motions, learned from these examples. In one embodiment, the motion signals are divided up into "segments" of a fixed, short temporal length. The prior model is a probability distribution over those temporal segments of marker positions over a few frames.

If one chooses too many frames for units of human motion, the training data will not be long enough to give a reliable enough model of such a complex vector. If one chooses too few frames, one cannot capture enough motion regularities with the subject model. It has been found that 10 is a good number of frames for each segment. Sampling from the original data, with an overlapped offset of 5 frames, 257 10-frame segments are obtained, represented by 1110 numbers each, e.g. 37 body markers times 3 dimensions times 10 frames.

In order to obtain a probabilistic model that describes these vectors and motivated by the success of principle components analysis, PCA, at dimensionality reduction as described by R. O. Duda and P. E. Hart in Pattern classification and scene analysis, Wiley-Interscience, 1973. and M. Turk and A. Pentland. Eigenfaces for recognition in J. of Cognitive Neuroscience, 3(1), 1991., one first ascertains whether one can describe these motion segments as linear combinations of basis functions. This is accomplished by forming a training matrix, M, by stacking the 1110 dimensional training vectors together in columns, after first subtracting the mean vector, m. Singular value decomposition, SVD, gives M=USV', where the columns of U are the basis functions and the diagonal elements of S are the corresponding singluar values. The solid line of FIG. 4 shows the singular value spectrum. The spectrum drops quickly, allowing a good summary of the data, e.g. 91% of the variance, from just 50 eigenvectors.

FIG. 3 shows a typical motion sequence synthesized using 40 eigenvectors, showing an imperfect, but good, reconstruction. Thus, one can summarize the 1110 dimensional motion segment vectors by their coordinates in a 50 dimensional subspace of the 1110 dimensional space.

Of course, the singular values themselves provide additional information about the motion segment data. One can model the data as resulting from a gaussian probability distribution of covariance A=US as described above in the aforementioned book by Duda. This probabilistic model is much stronger than just the subspace information itself as described by B. Moghaddam and A. Pentland, Probabilistic visual learning for object detection, Proc. 5th Intl. Conf. Computer Vision, pages 786-793. IEEE, 1995.

FIG. 5 shows three random draws from the resulting probability model for 10-frame human motion segments. The motions look like plausible human motions, although, of course, random. This gaussian distribution provides a useful prior model for how a human moves over time. Yet, as will be seen, it is simple enough to work with easily and to provide some analytic estimation results.

The goal is to infer 3-dimensional human motion from an image sequence. In order to do this one utilizes simplified rendering conditions, with the simplified rendering conditions being as follows: the body is transparent, and each marker is rendered to the image plane orthographically. For figural motion described by human motion basis coefficients .alpha., the rendered image sequence, y, is:

y=PU.alpha., (2) where P is the projection operator which collapses the y dimension of the image sequence Ua. Note that under these rendering conditions, the markers are distinguishable from each other.

To estimate the figure's 3-d motion, one wants to find the most probable 3-d explanation, specified by .alpha., for a given 2-d observation of markers over time, y. By Bayes theorem, one has

P(.alpha..backslash.y)=k.sub.1 P(y.backslash..alpha.)P(.alpha.),(3) where k.sub.1 is a normalization constant independent of the parameters .alpha. that one seeks to optimize. As developed above, for the prior probability, P(.alpha.), the multidimensional gaussian is P(.alpha.)=k.sub.2 e.sup.-.alpha..LAMBDA..spsp.-1.sub..alpha.,(4) where k.sub.2 is another normalization constant. If one models the observation noise as i.i.d. gaussian with variance .sigma., one has, for the likelihood term of Bayes theorem, P(y.backslash..alpha.)=k.sub.e e.sup.-.backslash.y-PU.alpha..backslash..spsp.2.sub./(2.sigma..spsp.2),(5) with normalization constant k.sub.3.

The posterior distribution is the product of these two gaussians. That yields another gaussian, with mean and covariance found by a matrix generalization of "completing the square", as described by A. Gelb, editor, Applied optimal estimation, MIT Press, 1974. The squared error optimal estimate for .alpha. is then

.alpha.=SU'P'(PUSU'P'+.sigma.I).sup.-1 (y-(Pm)) (6)

FIGS. 6a-d illustrate applying this estimate to an overlapped sequence of 3 motion segments of 20 frames, each 10 frame segment offset by 5 frames. In one experiment one of the 10 sequences was omitted from the training data, and a subset of it was used for this test. FIG. 6a shows the original sequence, and FIG. 6b, the orthographic projection. FIG. 6c is the 3-d reconstruction resulting from the likelihood term alone, omitting the gaussian prior information. This finds the coefficients of the human motion basis functions which best explain the visual data. Note that the 3-d reconstruction is poor. FIG. 6d is the full Bayesian solution of Eq. 6. Note including the prior information gives a much better 3-d reconstruction. The subject gaussian probability model and the simplified rendering conditions allow this analytic solution for the optimal 3-d motion estimate.

One also knows the covariance matrix, Q, describing the uncertainty in the estimated 3-d configuration after viewing the 2-d sequence,

Q=S-SU'P'(PUSU'P'+.sigma.I).sup.-1 PUS, (7) where I is the identity matrix, the rows and columns having the dimensionality of the observations.

Of course, without any human motion model, the depth of each marker would be completely unknown. The prior model for human motion of the subject system removes most of those ambiguities. The structure of the posterior covariance Q reveals what ambiguities that remain in the 3-d structure after viewing the image sequence. FIG. 4 shows a comparison of the diagonal terms of the prior covariance, e.g. the solid line, with those of the posterior covariance, e.g. the dashed line. One mode, mode 2, shows virtually no reduction in uncertainty. This corresponds to a rigid translation mode moving nearly along the line of sight of the camera, as shown in FIG. 7a. The second highest uncertainty mode, mode 1, is another rigid translation mode, shown in FIG. 7b. The non-rigid mode having the highest posterior uncertainty, mode 10, is shown in FIG. 7c. This mode spreads the individual's arms along the line of sight of the camera. Note that this high uncertainty mode reflects the errors observed in the reconstruction of FIG. 6d.

Under the subject gaussian prior model, one can quantify how much 3-d motion information one gains from seeing the orthographic projections of the marker positions over time. For the example of FIGS. 6a-d, the prior probability distribution occupies a certain volume in the 50 dimensional parameter space. The ratio of that volume to the posterior probability distribution's volume is 10.sup.-14. While it is hard to gauge high dimensional volumes intuitively, the ratio is small and the posterior uncertainty is considerably reduced. In post-processing, one might expect to also remove the rigid mode ambiguities, either by user interaction, or by applying rigid ground contact constraints.

The following conclusions can be drawn from studying the problem in this simplified rendering domain. Using prior knowledge of human motion indeed does improve the 3-d reconstructions possible from monocular image sequence data. For the subject training set of human motions, the remaining uncertainty after observations lay in the rigid translation away from the camera, and in a mode spreading the arms along the camera ray. The reconstructions are generally good. The image information used is the 2-d projections of marker positions of a stick figure. The conclusion is that if one were able to accurately overlay a 2-d stick figure on top of the human figure in a video sequence, approximately orthrographically rendered, one should be able to achieve comparable 3-d reconstruction accuracy from real images.

In order to estimate the 3-d body motion, one first wants to find a stick figure summary of the 2-d moving image of a human figure. This is a problem that various research groups have addressed, and, to a large degree, solved, as described by Hager and Belheumer, and Black. Moreover collaborators have developed parameterized motion models for tracking particular human actions as described by G. D. Hager and P. N. Belhumeur, Real-time tracking of image regions with changes in geometry and illumination, Proc. IEEE CVPR, pages 403-410, 1996 and M. J. Black, Y. Yacoob, A. D. Jepson, and D. J. Fleet, Learning paramterized models of image motion, Proc. IEEE CVPR, pages 561-567, 1997. In an article by A. Blake and M. Isard entitled 3D position, attitude and shape input using video tracking of hands and lips, Proc. SIGGRAPH 94, pages 185-192, 1994, In Computer Graphics, Annual Conference Series, these researchers have developed contour-based tracking of non-rigid objects. Note also Pfinder by C. Wren, A. Azarbayejani, T. Darrell, and A. Pentland, Pfinder: real-time tracking of the human body, Intl. Conf. on Automatic Face and Gesture Recognition, pages 51-56, Killington, Vermont, 1996, IEEE Computer Society, tracks the human figure over stationary environments.

The following desribes the subject tracking method. Because it was important to reconstruct 3-d figures even from difficult film sequences, interactive correction of the mistakes of the automatic tracking were allowed. This is in addition to the other interactions needed for the problem: to specify which human to track, and over which time period.

The goal was to demonstrate that the reconstruction method developed for the simplified rendering carried over to natural scenes. The image rendering, over the time frames of interest, was assumed to be roughly orthographic, with the effect of a moving background being ignored. The figure's motion relative to the camera frame was reconstructed, not the background.

To achieve some independence of clothes patterns of the human figure, the gradient of the image intensities was taken, and the gradient strengths were normalized by a blurred average of the local contrast strength. These normalized edge strengths were then blurred enough to make a low-resolution sampling of 10 by 8 sensors.

Based on the location of the sticks of the stick figure, a prediction for what the sensors ought to see was formed, assigning a fixed edge strength to each stick. The squared difference between the observed sensor responses and the predictions was penalized.

The user can interactively specify the correct location of any stick figure part at any time frame. This effectively places a spring between the image position and the stick figure part at that particular time.

These two inputs are integrated with the prior information in an function optimization scheme. The .alpha. being sought is that which minimizes an energy, E(.alpha.), from the image data. The symbol .LAMBDA..sup.-1 is the inverse of the Gaussian probability distribution of the covariance according to Duda as described above. ##EQU2## R is the vector of sensor responses over time from the image data. The function f converts .alpha. body motion coefficients to predicted sensor responses. I.sub.i is the ith point position specified by the user, and P.sub.i projects the .alpha. coefficients onto the corresponding ith stick figure part 2-d position. .lambda..sub.1 and .lambda..sub.2 are constants which reflect the weights of the image data, the priors over human motions, and the interactively specified 2-d point matches.

In the Bayesian framework, E is interpreted as the negative log of the posterior probability. .lambda..sub.1 and .lambda..sub.2 then represent observation and user "noise strengths". The quadratic penalty for sensor response differences is the log of the likelihood term, and both the interactively placed springs and the gaussian prior motion model represent prior information about the parameters. Also included is a 90.degree. rotated version of all training data in the calculation of the prior probability. The 3-d body motion parameters are found which maximize the posterior probability.

The recovered optimal .alpha. yields the recovered marker positions over time. Those marker positions are then fit to cylinder positions in a simple figure model using least squares techniques.

FIGS. 8a-c show the resulting estimates of the 3-d figure positions from a 100 frame sequence of Barishnikov dancing. In order to test the 3-d reconstruction algorithm, rather than the 2-d tracking algorithm, approximately one interactive location specification per frame were used to ensure the accurate overlay of a stick figure over the motion sequence. FIG. 8a shows the input sequence and the overlaid stick figure.

E in Eq. 1 is minimized to find the .alpha. estimate over each overlapped 10-frame segment. Positions from the overlapped segments were linearly blended together. The offset away from the camera of each segment was set to ensure continuity to the next segment.

FIG. 8b shows the recovered 3-d marker positions, viewed from 30.degree. away from the camera position. FIG. 8c shows the 3-d cylinder model, viewed from that same off-camera position. Given the simple gaussian prior model for human motion, the results are remarkably strong. The 3-d dance is rendered as a linear combination of basis motions learned from the motion capture training set. The dancing cylinder figure generally captures the 3-d motions of the dancer. The cylinder dancer does not point his toes, as Barkshnikov does, but toe pointing was not in the training set.

A statistical model was used to infer the 3-d positions of a human figure from an image sequence of the human moving. The model of human motion was learned from a training set of 3-d motion capture data, obtained from a calibrated tracking system.

Those results show how to estimate the 3-d figure motion if one can place a 2-d stick figure over the image of the moving person. Such a tracker was developed, allowing interactive correction of tracking mistakes, to test the 3-d recovery method. Good recovery of 3-d motion for a difficult dance sequence, viewed from a single camera, was obtained. These results show the power of adding prior knowledge about human motions, in a Bayesian framework, to the problem of interpreting images of people.

Note that there are three improvements which make the 3-d estimation more robust. As a first improvement, the first term in the sum of Equation 8 can be replaced by ##EQU3## where t is an index of the frame time of each frame of the motion segment. I.sub.t is a vector of the image intensities at time t at the 2-d position of its vector argument. M(.alpha.) is the 2-d positions of each marker for the human motion represented by the linear combination of training segments given by the combination coefficients .alpha.. Thus it can be seen that the coefficients .alpha. which minimize this term will correspond to a figure motion in which the image intensities at the marker positions in the image change the least from frame to frame. This will encourage the figure model to track the human figure in the video.

Secondly, the motion segments can be estimated sequentially, and are often overlapped in time. For example, motion segment 1 may correspond to video frames 1 through 10; motion segment 2 may correspond to video frames 2 through 11, etc. For each segment, a 3-d estimate is formed. The 3-d reconstruction for a particular video frame number is computed as the average of estimates of all the overlapping motion segments.

Thus, it can be seen that prior to finding the optimium .alpha. for Equation 8, an initial estimate of the marker positions for various time frames can be calculated from the motion estimates from the previous motion segments.

It will be apparent that an appropriate subset of the training segments may be used in the linear combination to be optimized. Using all the training segments at all times can lead to too many degrees of freedom for the model and in poor reconstructions. Reconstructions can be improved by restricting the segments used in the optimization to a small number, say 5 or 10, which are close to the current best estimate of the current snippet.

The second term can be replaced by .lambda..sub.1 ((.alpha.-.alpha.)'.LAMBDA..sup.-1 (.alpha.-.alpha.), where .alpha. is the prior mean of the coefficients .alpha.. In a preferred embodiment, .alpha.) is set to the average of the small number of selected training segments, ##EQU4## for each of the N selected training segments.

A program listing describing the operation of the subject system is contained in the appendix hereto, with the program being written in Matlab.

__________________________________________________________________________file: /homes/leventon/Matlab/G2a.mfunction v = G2a(x, y)v = 0.9213*(2*x*x-)*exp(-(x*x+y*y));file: /homes/leventon/Matlab/G2b.mfunction v = G2b(x, y)v = (0.9213*2)*x*y*exp(-(x*x+y*y));file: /homes/leventon/Matlab/G2c.mfunction v = G2c(x, y)v = 0.9213*(2*y*y-1)*sxp(-x*x+y*y));file: /homes/leventon/Matlab/G2theta.mfunction v = G2theta(x, y, theta)v = cos(theta)2* G2a(x, y) - 2*cos(theta)*sin(theta) * G2b(x, y) +sin(theta) 2 * G2c(x, y);file: /homes/leventon/Matlab/H2a.mfunction v = H2a(x, y)v = 0.9780*(-2.254*x+x 3)*exp(-(x*x+y*y));file: /homes/leventon/Matlab/H2b.mfunction v = H2d(x, y)v = 0.9780*(-0.7515+x 2)*y*exp(-(x*x+y*y));file: /homes/leventon/Matlab/H2c.mfunction v = H2d(x, y)v = 0.9780*(-0.7515+y 2)*x*exp(-(x*x+y*y));file: /homes/leventon/Matlab/H2d.mfunction v = H2d(x, y)v = 0.9780*(-2.254*y+y 3)*exp(-(x*x+y*y));file: /homes/leventon/Matlab/H2theta.mfunction v = H2theta(x, y, theta)v = cos(theta) 3 * H2a(x, y);v = v - 3*cos(theta) 2*sin(theta) * H2b(x,y);v = v + 3*cos(theta)*sin(theta) 2 * H2c(x,y);v = v - sin(theta) 3 * H2d(x, y);file: /homes/leventon/Matlab/H4a.mfunction x = H4a(x, y)x = 0.3975 * (7.189*x - 7.501*x 3 + x 5)*exp(-(x 2+y 2));file: /homes/leventon/Matlab/aToR.mfunction [R, Rx, Ry, Rz] = aToR(a);c = cos(-a);s = sin(-a)Rx = [1 0 0; 0 c(1) s(1); 0 -s(1) c(1)];Ry = [c(2) 0 s(2); 0 1 0; -s(2) 0 c(2)];Rz = [c(3) s(3) 0; -s(3) c(3) 0; 0 0 1];% R = Rz*Ry*Rx;R = Rx*Ry*Rz;file: /homes/leventon/Matlab/addConstraints.mfunction [ptcor, trackcor, itering, t, times.sub.-- changed] =addConstraints(seq, seqIm, ptcor, trackcor, t)% [ptcor, trackcor, itering, t] = addConstraints(seq, seqIm)% Starting with no constraints, the user interactively adds point or%% tracking constraints. Seq is the reconstructed 3d sequence.% seqIm is the Image sequence. This function returns the% correspondences in two cells, "ptcor" for the point% correspondences,% and "trackcor" for the tracking correspondences.% The return val itering tells if the user wants to continue% interating%, through the svd-loop. (Whether the user clicked "Done" or "Cancel"%% and whether the user added anything from the last iteration).%% The return val t stores what time step the user left off at.%% [ptcor, trackcor, itering, t] = addConstraints(seq, seqIm, ptcor,%, trackcor, t)% The ptcor and trackcor inputs allow for constraints to be% appended.% The t input deterinines at what time frame the interaction starts.% global ptcor% This function has pretty confusing control structure, but% matlab doesn't have the "continue" command . . .if (exists(seq, `scale`) == 0), seq.scale = 1; endif (exists(seqIm, `scale`) == 0), seqIm.scale = 1; endlinecolor = `b`; - scale = seqIm.scale / seq.scale;%% Display some instructionsdisp(`First Click:`);disp(`Use Left Mouse Button on any window to select marker`);disp(`Use Middle Button to step backward in time`);disp(`Use Right Button to step forward in time`);disp(` `);disp(`Second Click:`);disp(`Use Left Mouse Button on image window to selectcorrespondence`);disp(`Use Right Button on image window to define a trackingrectangle`);disp(` (first click upper left, then click lower right)`);disp(` Use Middle Mouse Button to CANCEL the first click`);if (nargin < 5) t = 1; endif (isempty(t)) t = 1; endn = size(seq.data, 1);d = seq.data(2:n, :);nFr = seq.nframes;global seqINFOnmkr = seqINFO.nummkr(seq.subtype);d = reshape(d', 3, nmkr, nFr);d = transpose2(d);if (nargin < 4) trackcor = cell(0); endif (nargin < 3) ptc9r = cell(0); end%% Dray Guy #1if (seq.subtype == 1)seqDiag = seqCrop(seqGet(6), 9, 9);elseif (seq.subtype == 2)seqDiag = seqCrop(seqGet(51), 9, 9);endguy1seq = seqConvert(seqDiag, `2d`);guy1dat = guy1seq.data(2:nmkr+1, :);guy2seq = seqConvert(seqRotate(seqDiag, pi/2), `2d`);guy2dat = guy2seq.data(2:nmkr+1, :);%% Dray Guy #1subplot(2,2,3);seqView(guy1seq);guy1.sub.-- figure = gca;axis off%% Dray Guy #2subplot(2,2,4);seqView(guy2seq);guy2.sub.-- figure = gca;axis off%% Start off by "showing" the already-established linksshowing = 1;%% We haven't made any changes yet . . .itering = 0;%% (There's no "do while" loop, so . . .)notdone = 1;%% There's no continue (only break), so . . .continue = 0;%% The last things we added (0-nothing, 1-ptcor, 2-trackcor) . . .lastwas = [ ];%% Times changedtimes.sub.-- changed = zeros(1,nFr);while (notdone)subplot(2,1,1);hold off%% Overlay the sequence on the image at time tseqView2({seqIm, seq}, t);axis off%% Add our "Buttons" (hee hee)H1 = text(scale*0, scale*(-5), `Done!`);H2 = text(scale*40, scale*(-5), `Cancel`);if (showing)H3 = text(scale*80, scale*(-5), `Hide`);elseH3 = text(scale*80, scale*(-5), `Show`);endHt = text(scale*120, scale*(-5), sprintf(`Time = *%d`, t));H4 = text(scale*160, scale*(-5), `Undo`);main.sub.-- figure = gca;% Transfer the 3d data from world to camera coordinatesd2 = worldTocam(d(:,:,t));if (showing)%% Display all the Point Correspondencesfor i = 1:length(ptcor)if (ptcor{i}.t == t)mkr = ptcor{i}.mkr;p2 = ptcor{i}.pos([2 1]);subplot(2,1,1);hold onplot(scale*[d2(mkr, 1) p2(1)], scale*[d2(mkr, 2) p2(2)], . . .[linecolor `o`]);endend%% Display all the Tracking Correspondencesfor i = i:iength(trackcor)if (trackcor{i}.t == t)mkr = trackcor{i}.mkr;pul = trackcor{i}.pos(i,[2 1]);plr = trackcor{i}.pos(2,[2 1]);p2 = (pul+plr)/2;subplot(2,1,1);hold onplot(scale*[d2(mkr, 1) p2(1)], scale*[d2(mkr, 2) p2(2)], . . .[linecolor `o-`]);if (trackcor{i}.first == t) plot(scale*[pul(1) pul(1) plr(1) plr(1) pu1(1)],scale*[pul(2) plr(2) plr(2) pul(2) pul(2)], `b+-`);else plot(scale*[pul(1) pul(1) plr(1) plr(1) pul(1)],scale*[pul(2) plr(2) plr(2) pul(2) pul(2)], `g+-`);endendendend%% "Matlab's control structure" . . . "sucks?" (said like "LSC . . .")if (continue == 0)[x, y, b] = ginput(1);win = gca;p = [x y];if (win==main.sub.-- figure), p=p/scale; endendcontinue = 0;%% While we haven't ciicked to the left of the imagewhile (p(2) > 0)%% Check which button we clicked . . . adjust time if necessaryif (b == 2)t = t - 1;if (t < 1)t = 1;endbreakelseif (b == 3)t = t + 1;if (t > nFr)t = nFr;endbreakend%% Find WHICH led we just clickedif (win == main figure)[v, which] = min(sum((d2-ones(nmkr,1)*p). 2, 2));elseif (win == guy1.sub.-- figure)[v, which] = min(sum((guy1dat-ones(nmkr,1)*p). 2, 2));elseif (win == guy2.sub.-- figure)[v, which] = min(sum((guy2dat-ones(nmkr,1)*p). 2, 2));elsebreakendhold on%% Highiight the led that was clickedsubplot(2,1,1);plot(scale*d2(which, 1), scale*d2(which, 2), [linecolor `*`]);subplot(2,2,3);seqView(guy1seq, [ ], cons(`viewang`, [0 0], `leds`, which-1));axis offsubplot (2,2,4);seqView(guy2seq, [ ], cons(`viewang`, [-90 0], `leds`, which-1));axis off%% Get the "New" position[x, y, b] =ginput(1);win = gca;pnew = [x y];if (win==main.sub.-- figure), pnew=pnew/scale; endif (b == 2)%% Middle Button -- cancel!break;endif (win .sup..about. = main.sub.-- figure)p = pnew;continue = 1;breakendif ((x < 0) .parallel. (y < 0))breakendif (b == 3)%% Right button -- second click means define a trackingrectanglepul = pnew;subplot(2,1,1);plot(scale*d2(which, 1), scale*d2(which, 2), `b*`);plot(scale*pul(1), scale*pul(2), `b+`);[x, y, b] = ginput(1);win = gca;plr = [x y];if (win==main.sub.-- figure), plr=plr/scale; endif (win .sup..about. = main.sub.-- figure)breakendplot(scale*[pul(1) pul(1) plr(1) plr(1) pul(1)], scale*[pul(2)plr(2) plr(2) pul(2) pul(2)], `b+-`);pnew = (pul+plr)/2;ntcor = length(trackcor)+1;times.sub.-- changed(t) = 1;trackcor{ntcor}.mkr = which;trackcor{ntcor}.t = t;trackcor{ntcor}.first = t;trackcor{ntcor}.pos = [pul; plr];trackcor{ntcor}.pos = trackcor{ntcor}.pos(:, [2 1]);trackcor{ntcor}.postrackcor{ntcor}.cert = 1;trackcor{ntcor}.hausfrac = 1;trackcor{ntcor}.weight = 1;lastwas = [2 lastwas];else%% Left button -- second click means define a pointnptcor = length(ptcor)+1;times.sub.-- changed(t) = 1;ptcor{nptcor}.mkr = which;ptcor{nptcor}.t = t;ptcor{nptcor}.first = t;ptcor{nptcor}.pos = pnew([2 1]);ptcor{nptcor}.cert = 1;ptcor{nptcor}.hausfrac = 1;ptcor{nptcor}.weight = 1;lastwas = [1 lastwas];endsubplot (2,1,1);plot(scale*[d2(which, 1) pnew(1)], scale*[d2(which, 2) pnew(2)],. . .[linecolor `o-`]);%% We changed something . . . Keep on minimizing when "done" isclickeditering = 1;[x, y, b] = ginput(1);win = gca;p = [x y];if (win==main.sub.-- figure), p=p/scale; endendif (p(2) < 0)if (p(1) < 75)%% Either "Done" or "Cancel" was clickednotdone = 0;elseif (p(1) < 115)%% Show/Hide was Clickedshowing = .about.showing;else%% Undo was Clickedif (isempty(lastwas) == 0) if (lastwas(1) == 1) ptcor = ptcor(1:length(ptcor)-1); elseif (lastwas(1) == 2) trackcor = trackcor(1:length(trackcor)-1); end lastwas = lastwas(2:length(lastwas));endendendendif (p(1) > 35)%% They Clicked "CANCEL" (not "Done")%% So let the caller know . . .itering = 0;enddelete(H1);delete(H2);delete(H3);delete(H4);if (isempty(lastwas))for i = 1:length(ptcor)times.sub.-- changed(ptcor{i}.t) = 1;endfor i = 1:length(trackcor)times.sub.-- changed(trackcor{i}.t) = 1;endendfile: /homes/leventon/Matlab/aencode.mfunction M = aencode(varargin)params = varargin{nargin};n = nargin - 1;s = params(1);k = params(2);M = [ ];c = 1;A = varargin{1};ndeg = 28;nRows = ndeg * s;for w = 1:nA = varargin{w};nfr = size(A, 2);for i = 1:k:(nfr-s-1)v = vectorize(A(1:ndeg, i:i+s-1)`);M(1:nRows,c) = v;c = c + 1;endendfile: /homes/leventon/Matlab/aencode2.mfunction M = aencode2(varargin)params = varargin{nargin};n = nargin - 1;s = params(1);k = params(2);M = [ ];c = 1;A = varargin{i};ndeg = 20;nRows = ndeg * s;for w = 1:nA = varargin{w};nfr = size(A, 2);for i = i:k:(nfr-s+1)v = vectorize(A(4:23, i:i+s-1)`);M(i:nRows,c) = v;c = c + 1;endendfile: /homes/leventon/Matlab/allbut.mfunction [Xnx, Mnx, unx, snx, vnx, En] = allbut(ALL, x, nEv)ISx = ALL{x};N = size(ALL, 1);if (x == 1)ISnx = {ALL{2:N}};elseif (x == N)ISnx = {ALL{i:N-1}};elseISnx = {ALL{1:x-1}, ALL{x+1:N}};endMnx = mencode(ISnx{1:N-1}, [10 5]);[unx, snx, vnx, Enx] = decomp(Mnx, 37);Xnx = project(ISx, unx, nEv);file: /homes/leventon/Matlab/allbut23.mfunction [A23nxp, A2nxp, M2nx, M3nx, u2nx, s2nx, v2nx, E2nx] =allbut23(ALL2, ALL3, x, r, nEv)[N, M] = size(ALL2);M = 3;IS2x = ALL2{x, r};IS2nx1 = cell(1, M);IS3nx1 = cell(1, M);for m = 1:Mif (x == 1)IS2nx1{m} = {ALL2{2:N,m}}`;IS3nx1{m} = {ALL3{2:N,m}}`;elseif (x == N)IS2nx1{m} = {ALL2{1:N-1,m}}`;IS3nx1{m} = {ALL3{1:N-1,m}}`;elseIS2nx1{m} = {ALL2{1:x-1,m}, ALL2{x+1:N,m}}`;IS3nx1{m} = {ALL3{1:x-1,m}, ALL3{x+1:N,m}}`;endendIS2nx = cat(2, IS2nx1{1:M});IS3nx = cat(2, IS3nx1{1:M});M2nx = mencode2(IS2nx{1:N-1,1:M}, [10 5]);M3nx = mencode(IS3nx{1:N-1,1:M}, [10 5]);[u2nx, s2nx, v2nx, E2nx] = decomp2(M2nx, 37);[A23nxp, A2nxp] = project23(IS2x, M3nx, u2nx, s2nx, v2nx, nEv);file: /homes/leventon/Matlab/anglesc.mMaw = aencode(walkloopa, [10 5]);[u, s, v] = svd(Maw);viewang(person, matricize(Maw(;,1), 23)`);viewang(person, matricize(u(:,1), 23)`);Mp = u(:,1:100) * u(:,1:100)`* Maw(:,1); viewang(person,matricize(Mp, 23)`);file: /homes/leventon/Matlab/aoaNegate.mfunction M2 = aoaNegate(M)[n, lose] = size(M);M2 = [M(1,:); -M(2:n,:)];file: /homes/leventon/Matlab/aoaerr2.mfunction err = aoaerr2(A1, A2)nfr1 = A1(1,1);nfr2 = A2(1,1);n= 37 * min([nfri nfr2]);B1 = A1(2:n+1, :);B2 = A2(2:n+1, :);err = 0;for i = i:nerr = err + (sum((B1(i,:-B2(i,:)). 2));enderr = sqrt(err/n);file: /homes/leventon/Matlab/aoaerr23 .mfunction err = aoaerr23(A1, A2)nfr1 = A1(1,1;nfr2 = A2(1,1);t = min([nfr1 nfr2]);n = 37 * t;B1 = A1(2:n+1, :);B2 = A2(2:n+1, :);err = 0;for i = 1:tr = ((i-1)*37+1):i*37;com1 = (sum(B1(r,:))/37);com2 = (sum(B2(r,:))/37);for j = 1:37err = err + sum(((B1(r(j),:)-com1)-(B2(r(j),:)-com2)). 2);endenderr = sqrt(err/n);file: /homes/leventon/Matlab/aoaerr3.mfunction err = aoaerr3(A1, A2)nfr1 = A1(1,1);nfr2 = A2(1,1);n = 37 * min([nfr1 nfr2]);B1 = A1(2:n+1, :);B2 = A2(2:n+1, :);err = 0;for i = 1:nerr = err + (sum((B1(i,:)-B2(i,:)). 2));enderr = sqrt(err/n);file: /homes/leventon/Matlab/args.mfunction a = cons(varargin)n = length(varargin);if (n==0)a.struct = [ ];returnendif (mod(n, 2) == 0)a = struct(varargin{:});elsea = varargin{1};f = fieldnames(a);nf = size(f, 1);for i=2:2:nfound = 0;for j = 1:nfif (strcmp(varargin{i}, f(j))) found = 1;endendif (found == 0)a = setfield(a, varargin{i}, varargin{i+1});endendendfile: /homes/leventonlMatlab/atan2e.mfunction t = atan2e(a, b, e)if (abs(a) < e)a = 0;endif (abs(b) < e)b = 0; -endt = atan2(a, b);file: /homes/leventon/Matiab/autocrop.mfunction [I, amount] = autocrop(I, val)if (nargin < 2)Ib = [I(:,1)' I(1,:) I(:.size(I,2))' I(size(I,1),:)];val = mode(Ib(:));endJ = (I .sup..about. = val);Jh = find(sum(J,1));Jv = find(sum(J,2));if (isempty(Jh))I = [ ]; - returnendbx = Jv(1);ex = Jv(length(Jv));by = Jh(1);ey = Jh(length(Jh));amount = [(bx-1) (by-1) (size(I,i)-ex) (size(I,2)-ey)];I = I(bx:ex, by:ey);file: /homes/leventon/Matlab/autocropr.mfunction [rx, ry] = autocropr(I, val)if (nargin < 2)Ib = [I(:,1)'I(1,:) I(:,size(I,2))' I(size(I,1),:)];val = mode(Ib(:));endJ = (I .sup..about. = val);Jh = find(sum(J,1));Jv = find(sum(J,2));if (isempty(Jh))rx = [ ];ry = [ ];returnendbx = Jv(1);ex = Jv(length(Jv));by = Jh(1);ey = Jh(length(Jh));rx = bx:ex;ry = by:ey;if (nargout < 2)rx = {rx, ry};file: /homes/leventon/Matlab/barish.mseqInitload /homes/leventon/Data/barish.matSEQUENCES = 1:10;VIEW.sub.-- ANGLES = (0:7)*pi/4;sel = seqSelect(SEQUENCSE, VIEW.sub.-- ANGLES);knlPulse = seqTrain(sel, `pulse`, `submean`, 1, `MtM`, 0, `headalign`,1);barish =seqRecon(file: /homes/leventon/Matlab/block.mfunction MM = block(M, k)[m, n] = size(M);MM = zeros(k*m, k*n);for i=i:kMM((i-1)*m+1:i*m, (i-1)*n+1:i*n) = M;endfile: /homes/leventon/Matlab/brent.m%% brent.m function implements brent's method for 1-d%% line minimization. see p. 404, num rec. in C.%% Feb. 3, 1997 freeman typed in and debugged. seems ok.function [fx, xmin] = brent(ax, bx, cx, FUN, tol)evalstr = [FUN];ITMAX = 100;CGOLD = 0.3819660;ZEPS = 1.0e-10;e = 0.0;if (ax < cx); a = ax; else a = cx; end;if (ax > cx); b = ax; else b = cx; end;x = bx; w = bx; v = bx;fx = eval([evalstr, `(x)`]); fv = fx ; fw = fx;for iter = 1:ITMAXxm = 0.5 * (a + b);toll = tol * abs(x) + ZEPS;tol2 = 2.0 * toll;if (abs(x-xm) <= (tol2 - 0.5*(b-a)))xmin = x;return;endif (abs(e) > toll)r = (x-w)*(fx-fv);q = (x-v)*(fx-fw);p = (x-v)*q-(x-w)*r;q = 2.0*(q-r);if (q > 0.0) p = -p; end;q = abs(q);etemp = e;e = d;if ( (abs(p) >= abs(0.5*q*etemp)) .vertline. . . . (p <= q*(a-x)) .vertline. . . . (p >= q*(b-x)))if (x >= xm) e = a-x; else e = b-x; end;d = CGOLD*e;elsed = p/q;u = x+d;if ( (u-x < tol2) .vertline. (b-u < tol2) ) d = abs(toli)*sign(xm-x);end;endelseif (x >= xm) e = a-x; else e = b-x; end;d = CGOLD*e;endif (abs(d) >= toli)u = x+d;elseu = x+ abs(toli)*sign(d);endfu = eval([evalstr, `(u)`]);if (fu <= fx)if (u >= x) a = x; else b = x; end;v = w; w = x; x = u;fv = fw; fw = fx; fx = fu;elseif (u < x) a = u; else b = u; end;if ( (fu <= fw) .vertline. (w == x))v = w; w = u; fv = fw; fw = fu;elseif ( (fu <= fv) .vertline. (v == x) .vertline. (v == w) ) v = u; fv = fu;end;end;endenddisp(sprintf(`Too many iterations in brent`));xmin = x;return;file: /homes/leventon/Matlab/call.sub.-- canny.mfunction Ie = call.sub.-- canny(I, smth, canny1, canny2)SMOOTH.sub.-- CMD = `/homes/leventon/libgregk/bin/smooth`;CANNY.sub.-- CMD = `/homes/leventon/libgregk/bin/canny`;PEMTOPGM.sub.-- CMD = `/homes/leventon/libgregk/bin/pbmtopgm`;if (nargin < 2), smth = 1; endif (nargin < 3), canny1 = 2; endif (nargin < 4), canny2 = 5; endoutfn = sprintf(`tmp-%04.out`, round(mod(now, 10000)));infn = sprintf(`tmp-%04d.in`, round(mod(now, 10000)));if (size(I,3) == 3)I = 0.299 * I(:,:,1) + 0.587 * I(:,:,2) + 0.114 * I(:,:,3);endpgmWrite(I, outfn);cmd = sprintf(`%s-g %d < %s .vertline. %s -t %d %d .vertline. %s >! %s`,SMOOTH.sub.-- CMD,%smth, outfn, CANNY.sub.-- CMD, canny1, canny2, PBMTOPGM.sub.-- CMD,%infn);unix(cmd);Ie = pginRead(infn);cmd = sprintf(`/bin/rm -f %s %s`, outfn, infn);unix(cmd);file: /homes/leventon/Matlab/camToWorld.mfunction w = camToWorld(c)% camToWorld Take a list of points in camera coorinates,% And transform thein into world coordinates.w = zeros(size(c,1),3);w(:, 1) = (160-c(:, 1)-54)/(42-125)*(52.19+87.9)+43.19;w(:, 3) = (c(:, 2)-44)/(473-30)*(40-70)+40;file: /homes/leventon/Matlab/canny.mfunction [E, dx, dy] = canny(I, sd, th1, th0);%CANNY Edge detection.% E = canny(I) finds the edges in a gray scaled image I using% the Canny% method, and returns an image E where the edges of I are marked% by% non-zero intensity values.% E = canny(I, SD) uses SD as the standard deviation for the% gaussian% filtering phase. Default is 1 pixel.%% E = canny(I, SD, TH1) uses TH1 for the higher hysteresis% threshold.% Default is 0.5 times the strongest edge. Setting TH1 to zero% will% avoid the (sometimes time consuming) hysteresis.%% E = canny(I, SD, TH1, TH0) uses TH1 for the lower hysteresis% threshold.% Default is 0.1 times the strongest edge.%% See also EDGE (in the Image Processing toolbox).if nargin<2 sd= 1; end; if isempty(sd), sd= 1; end;if nargin<3 th1= .5; end; if isempty(th1), th1= .5; end;if nargin<4 th0= .1; end; if isempty(th0), th0= .1; end;x= -5*sd:sd*5;g= exp(-0.5/sd 2*x 2); % Create a normalized Gaussian%g= g(g>max(g)*.005); g= g/sum(g(:));dg= diff(g); % Gaussian first derivativedx= abs(conv2(I, dg, `same`)); % X/Y edgesdy= abs(conv2(I, dg', `same`));dx = smooth(dx, 1);dy = smooth(dy, 1);[ny, nx]= size(I); % Find maximady0= [dy(2:ny,:); dy(ny,:)]; dy2= [dy(1,:); dy(1:ny-1,:)];dx0= [dx(:, 2:nx) dx(:,nx)]; dx2= [dx(:,1) dx(:,1:nx-i)];peaks= find((dy>dy0 & dy>dy2) .vertline. (dx>dx0 & dx>dx2));peaks.sub.-- y = find((dy>dy0 & dy>dy2));peaks.sub.-- x = find((dx>dx0 & dx>dx2));e= zeros(size(I));e(peaks.sub.-- x)= dx(peaks.sub.-- x);e(peaks.sub.-- y) = max(e(peaks.sub.-- y), dy(peaks.sub.-- y));e(:,2) = zeros(ny,1); e(2,:)= zeros(1,nx); %, Remove artificial% edgese(: ,nx-2)= zeros(ny,1); e(ny-2, :)= zeros(1,nx);e(:,1) = zeros(ny,1); e(1,:)= zeros(1,nx);e(:,nx) = zeros(ny,1); e(ny,:)= zeros(1,nx);e(:,nx-)= zeros(ny,1); e(ny-1,:)= zeros(1,nx);e= e/max(e(:));if thi == 0, E= e; return; end % Perform hysteresisE(ny,nx)= 0;p= find(e >= thi);while length(p)E(p)= e(p);e(p)= zeros(size(p));n= [p+i p-i p+ny p-ny p-ny-1 p-ny+1 p+ny-1 p+ny+1]; % direct% neighborsOn= zeros(ny,nx); On(n)= n;p= find(e > tho & On);endfile: /homes/leventon/Matlab/ccm.mclear d;for i = 1:127;d{i}(: ,2) = (seq.data((i-1)*37+2:i*37+1, 2)-40)*(45.5-30)/(40-70) +43.5;d{i}(:,1) = 160-((seq.data((i-1)*37+2:i*37+1,i)-43.19)*(54-125)/(52.19+87.9) + 57);d{i} = [1 37; d{i}];endfile: /homes/leventon/Matlab/circle.mfunction c = circle(r, s)if (nargin < 2)s = r;endn = 2*r+1;c = zeros(2*s+1);offset = (s-r);for i = 1:nfor j = 1:nif (((i-i-r). 2+(j-1-r). 2) <= r. 2)c(i+offset,j+offset) = 1;endendendfile: /homes/leventon/Matlab/clickpts.mT = 94;step = 1;figure(1);colormap (gray)for i = 1:Timagesc(seqIm.data{i}(:,: ,2));title(sprintf(`%d`, i));head(i,:) = ginput(1);endfigure(1);colormap(gray)for i = 1:Timagesc(seqIm.data{i}(:,:,2));title(sprintf(`%d`, i));lefthand(i,:) = ginput(1);endfigure(1);colormap(gray)for i = 1:Timagesc(seqIm.data{i}(:,:,2));title(sprintf(`%d`, i));leftelbow(i,:) = ginput(1);endfigure(1);colormap(gray)for i = 1:Timagesc(seqIm.data{i}(:,:,2));title(sprintf(`%d`, i));leftknee(i,:) = ginput(1);endfigure(1);colormap(gray)for i = 1:Timagesc(seqIm.data{i}(:,:,2));title(sprintf(`%d`, i));lefthip(i,:) = ginput(1);endfigure(1);colormap(gray)for i = 1:Timagesc(seqIm.data{i}(:,:,2));title(sprintf(`%d`, i));leftfoot(i,:) = ginput(1);endfigure(1);colormap(gray)for i = 1:Timagesc(seqIm.data{i}(:,:,2));title(sprintf(`%d`, i));righthand(i,:) = ginput(1);endfigure(1);colormap(gray)for i = 1:Timagesc(seqIm.data{i}(:,:,2));title(sprintf(`%d`, i));rightelbow(i,:) = ginput(1);endfigure(1);colormap(gray)for i = 1:Timagesc(seqIm.data{i}(:,:,2));title(sprintf(`%d`, i));rightknee(i,:) = ginput(1);endfigure(1);colormap(gray)for i = 1:Timagesc(seqIm.data{i}(:,:,2));title(sprintf(`%d`, i));righthip(i,:) = ginput(1);endfigure(1);colormap(gray)for i = 1:Timagesc(seqIm.data{i}(:,:,2));title(sprintf(`%d`, i));rightfoot(i,:) = ginput(1);endT0 = 1;T = 101;step = 1;clear pt.sub.-- headi = 1;for t = [TO:step:T-1 T]pt.sub.-- head{i}.mkr = 20;pt.sub.-- head{i}.t = t;pt.sub.-- head{i}.first = 1;pt.sub.-- head{i}.cert = 1;pt.sub.-- head{i}.weight = 1;pt.sub.-- head{i}.hausfrac = 1;pt.sub.-- head{i}.pos = head(t, [2 1])/4;i = i + 1;endclear pt.sub.-- righthipi = 1;for t = [TO:step:T-1 T]pt.sub.-- righthip{i}.mkr = 8;pt.sub.-- righthip{i}.t = t;pt.sub.-- righthip{i}.first = 1;pt.sub.-- righthip{i}.cert = 1;pt.sub.-- righthip{i}.weight = 1;pt.sub.-- righthip{i}.hausfrac = 1;pt.sub.-- righthip{i}.pos = righthip(t,[2 1])/4;i = i + 1;endclear pt.sub.-- righthandi = 1;for t = [TO:step:T-1 T]pt.sub.-- righthand{i}.mkr = 36;pt.sub.-- righthand{i}.t = t;pt.sub.-- righthand{i}.first = 1;pt.sub.-- righthand{i}.cert = 1;pt.sub.-- righthand{i}.weight = 1;pt.sub.-- righthand{i}.hausfrac = 1;pt.sub.-- righthand{i}.Pos = righthand(t, [2 1])/4;i = 1 + 1;endclear pt.sub.-- rightfooti = 1;for t = [TO:step:T-1 T]pt.sub.-- rightfoot{i}.mkr = 26;pt.sub.-- rightfoot{i}.t = t;pt.sub.-- rightfoot{i}.first = 1;pt.sub.-- rightfoot{i}.cert = 1;pt.sub.-- rightfoot{i}.weight = 1;pt.sub.-- rightfoot{i}.hausfrac = 1;pt.sub.-- rightfoot{i}.pos = rightfoot(t, [2 1])/4;i = i + 1;endclear pt.sub.-- rightkneei = 1;for t = [TO:Step:T-1 T]pt.sub.-- rightknee{i}.mkr = 25;pt.sub.-- rightknee{i}.t = t;pt.sub.-- rightknee{i}.first = 1;pt.sub.-- rightknee{i}.cert = 1;pt.sub.-- rightknee{i}.weight = 1;pt.sub.-- rightknee{i}.hausfrac = 1;pt.sub.-- rightknee{i}.pos = rightknee(t, [2 1])/4;i = i + 1;endclear pt.sub.-- rightelbowi = 1;for t = [TO:step:T-1 T]pt.sub.-- rightelbow{i}.mkr = 17;pt.sub.-- rightelbow{i}.t = t;pt.sub.-- rightelbow{i}.first = 1;pt.sub.-- rightelbow{i}.cert = 1;pt.sub.-- rightelbow{i}.weight = 1;pt.sub.-- rightelbow{i}.hausfrac = 1;pt.sub.-- rightelbow{i}.pos = rightelbow(t, [2 1])/4,i = i + 1;endclear pt.sub.-- lefthipi = 1;for t = [TO:step:T-1 T]pt.sub.-- lefthip{i}.mkr = 5;pt.sub.-- lefthip{i}.t = t;pt.sub.-- lefthip{i}.first = 1;pt.sub.-- lefthip{i}.cert = 1;pt.sub.-- lefthip{i}.weight = 1;pt.sub.-- lefthip{i}.hausfrac = 1;pt.sub.-- lefthip{i}.pos = lefthip(t, [2 1])/4;i = i + 1;endclear pt.sub.-- leftfooti = 1;for t = [TO:step:T-1 T]pt.sub.-- leftfoot{i}.mkr = 23;pt.sub.-- leftfoot{i}.t = t;pt.sub.-- leftfoot{i}.first = 1;pt.sub.-- leftfoot{i}.cert = 1;pt.sub.-- leftfoot{i}.weight = 1;pt.sub.-- leftfoot{i}.hausfrac = 1;pt.sub.-- leftfoot{i}.pos = leftfoot(t, [2 1])/4;i = i + 1;endclear pt.sub.-- lefthandi = 1;for t = [TO:step:T-1 T]pt.sub.-- lefthand{i}.mkr = 31;pt.sub.-- lefthand{i}.t = t;pt.sub.-- lefthand{i}.first = 1;pt.sub.-- lefthand{i}.cert = 1;pt.sub.-- lefthand{i}.weight = 1;pt.sub.-- lefthand{i}.hausfrac = 1;pt.sub.-- lefthand{i}.pos = lefthand(t, [2 1])/4;i = i + 1;endclear pt.sub.-- leftkneei = 1;for t = [TO:step:T-1 T]pt.sub.-- leftknee{i}.mkr = 22;pt.sub.-- leftknee{i}.t = t;pt.sub.-- leftknee{i}.first = 1;pt.sub.-- leftknee{i}.cert = 1;pt.sub.-- leftknee{i}.weight = 1;pt.sub.-- leftknee{i}.hausfrac = 1;pt.sub.-- leftknee{i}.pos = leftknee(t, [2 1])/4;i = i + 1;endclear pt.sub.-- leftelbowi = 1;for t = [TO:step:T-1 T]pt.sub.-- leftelbow{i}.mkr = 13;pt.sub.-- leftelbow{i}.t = t;pt.sub.-- leftelbow{i}.first = 1;pt.sub.-- leftelbow{i}.cert = 1;pt.sub.-- leftelbow{i}.weight = 1;pt.sub.-- leftelbow{i}.hausfrac = 1;pt.sub.-- leftelbow{i}.pos = leftelbow(t, [2 1])/4;i = i + 1;endptcorr = cat(2, pt.sub.-- head, pt.sub.-- leftfoot, pt.sub.-- leftknee,pt.sub.-- lefthand,pt.sub.-- leftelbow, pt.sub.-- lefthip, pt.sub.-- rightfoot, pt.sub.--rightknee, pt.sub.-- righthand,pt.sub.-- rightelbow, pt.sub.-- righthip);%% Below is for subtype 2 (20 markers)TO = 1;T = 101;step = 1;clear pt.sub.-- headi = 1;for t = [TO:step:T-1 T]pt.sub.-- head{i}.mkr = 3;pt.sub.-- head{i}.t = t;pt.sub.-- head{i}.first = 1;pt.sub.-- head{i}.cert = 1;pt.sub.-- head{i}.weight = 1;pt.sub.-- head{i}.hausfrac = 1;pt.sub.-- head{i}.pos = head(t, [2 1])/4;i = i + 1;endclear pt.sub.-- righthipi = 1;for t = [TO:step:T-1 T]pt.sub.-- righthip{i}.mkr = 13;pt.sub.-- righthip{i}.t = t;pt.sub.-- righthip{i}.first = 1;pt.sub.-- righthip{i}.cert = 1;pt.sub.-- righthip{i}.weight = 1;pt.sub.-- righthip{i}.hausfrac = 1;pt.sub.-- righthip{i}.pos = righthip(t, [2 1])/4;i = i + 1;endclear pt.sub.-- righthandi = 1;for t = [TO:step:T-1 T]pt.sub.-- righthand{i}.mkr = 10;pt.sub.-- righthand{i}.t = t;pt.sub.-- righthand{i}.first = 1;pt.sub.-- righthand{i}.cert = 1;pt.sub.-- righthand{i}.weight = 1;pt.sub.-- righthand{i}.hausfrac = 1;pt.sub.-- righthand{i}.pos = righthand(t, [2 1])/4;i = i + 1;endclear pt.sub.-- rightfooti = 1;for t = [TO:step:T-1 T]pt.sub.-- rightfoot{i}.mkr = 17;pt.sub.-- rightfoot{i}.t = t;pt.sub.-- rightfoot{i}.first = 1;pt.sub.-- rightfoot{i}.cert = 1;pt.sub.-- rightfoot{i}.weight = 1;pt.sub.-- rightfoot{i}.hausfrac = 1;pt.sub.-- rightfoot{i}.pos = rightfoot(t, [2 1])/4;i = i + 1;endclear pt.sub.-- rightkneei = 1;for t = [TO:step:T-1 T]pt.sub.-- rightknee{i}.mkr = 15;pt.sub.-- rightknee{i}.t = t;pt.sub.-- rightknee{i}.first = 1;pt.sub.-- rightknee{i}.cert = 1;pt.sub.-- rightknee{i}.weight = 1;pt.sub.-- rightknee{i}.hausfrac = 1;pt.sub.-- rightknee{i}.pos = rightknee(t, [2 1])/4;i = i + 1;endclear pt.sub.-- rightelbowi = 1;for t = [TO:step:T-1 T]pt.sub.-- rightelbow{i}.mkr = 6;pt.sub.-- rightelbow{i}.t = t;pt.sub.-- rightelbow{i}.first = 1;pt.sub.-- rightelbow{i}.cert = 1;pt.sub.-- rightelbow{i}.weight = 1;pt.sub.-- rightelbow{i}.hausfrac = 1;pt.sub.-- rightelbow{i}.pos = rightelbow(t, [2 i])/4;i = i + 1;endclear pt.sub.-- lefthipi = 1;for t = [TO:step:T-1 T]pt.sub.-- lefthip{i}.mkr = 14;pt.sub.-- lefthip{i}.t = t;pt.sub.-- lefthip{i}.first = 1;pt.sub.-- lefthip{i}.cert = 1;pt.sub.-- lefthip{i}.weight = 1;pt.sub.-- lefthip{i}.hausfrac = 1;pt.sub.-- lefthip{i}.pos = lefthip(t, [2 i])/4;i = i + 1;endclear pt.sub.-- leftfooti = 1;for t = [TO:step:T-1 T]pt.sub.-- leftfoot{i}.mkr = 18;pt.sub.-- leftfoot{i}.t = 1;pt.sub.-- leftfoot{i}.first = 1;pt.sub.-- leftfoot{i}.cert = 1;pt.sub.-- leftfoot{i}.weight = 1;pt.sub.-- leftfoot{i}.hausfrac = 1;pt.sub.-- leftfoot{i}.pos = leftfoot(t, [2 1])/4;i = i + 1;endclear pt.sub.-- lefthandi = 1;for t = [TO:step:T-1 T]pt.sub.-- lefthand{i}.mkr = 11;pt.sub.-- lefthand{i}.t = t;pt.sub.-- lefthand{i}.first = 1;pt.sub.-- lefthand{i}.cert = 1;pt.sub.-- lefthand{i}.weight = 1;pt.sub.-- lefthand{i}.hausfrac = 1;pt.sub.-- lefthand{i}.pos = lefthand(t, [2 1])/4;i = i + 1;endclear pt.sub.-- leftkneei = 1;for t = [TO:step:T-1 T]pt.sub.-- leftknee{i}.mkr = 16;pt.sub.-- leftknee{i}.t = t;pt.sub.-- leftknee{i}.first = 1;pt.sub.-- leftknee{i}.cert = 1;pt.sub.-- leftknee{i}.weight = 1;pt.sub.-- leftknee{i}.hausfrac = 1;pt.sub.-- leftknee{i}.pos = leftknee(t, [2 1])/4;i = i + 1;endclear pt.sub.-- leftelbowi = 1;for t = [TO:step:T-1 T]pt.sub.-- leftelbow{i}.mkr = 7;pt.sub.-- leftelbow{i}.t = t;pt.sub.-- leftelbow{i}.first = 1;pt.sub.-- leftelbow{i}.cert = 1;pt.sub.-- leftelbow{i}.weight = 1;pt.sub.-- leftelbow{i}.hausfrac = 1;pt.sub.-- leftelbow{i}.pos = lefteibow(t, [2 1])/4;i = i + 1;endptcorr20 = cat(2, pt.sub.-- head, pt.sub.-- leftfoot, pt.sub.-- leftknee,pt.sub.-- lefthand,pt.sub.-- leftelbow, pt.sub.-- lefthip, pt.sub.-- rightfoot, pt.sub.--rightknee, pt.sub.-- righthand,pt.sub.-- rightelbow, pt.sub.-- righthip);file: /homes/leventon/Matlab/clip1Coeff.mfunction [xall, xsall] = clip1Coeff(encobs, knl, coeffs, nEv, lambdas,niter, lambda1)nSnip = size(coeffs, 2);%% The training matrix, MM = knl.enc;%% And it's singular value decompositionU = knl.U;V = knl.V;S = sparse(knl.S);%% I only want to use nEv eigenvectors. Zero the remaining eigen%% values.S(nEv+1:size(S,1), nEv+1:size(S,2)) = 0;S = sparse(S);s2 = diag(S.*S);%% For each snippet in the observationfor snip = 1:nSnipCLIPPING1.sub.-- SNIPPET = snip% Select out the initial guess of coefficients for this snippet% (computed by SVD)x0 = coeffs(:,snip);% Select out the observation for this snippety = encobs(:,snip);% Compute M'y ( M' = V*S'*U')% The reason only part of S is used is because less than% the total number of eigenvectors is used.Mty = V*S(1:size(S,2),1:size(S,2))`*U`*y;clear xs;lct = 1;x = x0;for lambda = lambdas% Compute inv(S+lambda*I)SpLi = sparse(size(S,2), size(S,2));SpLi(1:length(s2), 1:length(s2)) = diag(1./(s2 + lambda));SpLi = sparse(SpLi);MtMpLlpLIi = pinv(M'*M + lambda1 + lambda*eye(294));for ct = 1:niter%%% Step 2. Clip the coeffsxc = x .* (x > 0);%%% Step 3. Minimize%% E = (data - [training set]* coeffs) . 2 +%% lambda * (coeffs - clipped.sub.-- coeffs). 2%% See derivation, MEL's notes file "Friday, July 18th, 1997"% x = (V*SpLi*V') * (Mty + lambda*xc);x = MtMpLlpLIi * (Mty + lambda*xc + lambda1);xs{lct, ct} = x;endlct = lct + 1;endxall(:,snip) = x;xsall{snip} = xs;endfile: /homes/leventon/Matlab/clipCoeff.mfunction [xall, xsall] = clipCoeff(encobs, knl, coeffs, nEv, lambdas,niter)nSnip = size(coeffs, 2);%% The training matrix, MM = knl.enc;%% And it's singular value decompositionU = knl.U;V = knl.V;S = sparse(knl.S);%% I only want to use nEv eigenvectors. Zero the remaining eigen%% values.S(nEv+1:size(S,1), nEv+1:size(S,2)) = 0;S = sparse(S);s2 = diag(S.*S);%% For each snippet in the observationfor snip = 1:nSnipCLIPPING.sub.-- SNIPPET = snip% Select out the initial guess of coefficients for this snippet% (computed by SVD)x0 = coeffs(:,snip);% Select out the observation for this snippety = encobs(:,snip);% Compute M'y ( M' = V*S'*U')% The reason only part of S is used is because less than% the total number of eigenvectors is used.Mty = V*S(1:size(S,2),1:size(S,2)))`*U`*y;clear xs;lct = 1;x = x0;for lambda = lambdas% Compute inv(S+lambda*I)SpLi = sparse(size(S,2), size(S,2));SpLi(1:length(s2), 1:length(s2)) = diag(1./(s2 + lambda));SpLi = sparse(SpLi);for ct = 1:niter%%% Step 2. Clip the coeffs% Computing the clipping of the coefficientsxc = (x > 0) .* x;%%% Step 3. Minimize%% E = (data - [training set] * coeffs) . 2 +%% lambda * (coeffs - clipped.sub.-- coeffs) . 2%% See derivation, MEL's notes file "Friday, July 18th, 1997"x = (V*SpLi*V') * (Mty + lambda*xc);xs{lct, ct} = x;endlct = lct + 1;endxall(:,snip) = x;xsall{snip} = xs;endfile: /homes/leventon/Matlab/compConstr.mfunction [MtM, Mty, M.sub.-- s, y.sub.-- s, weight.sub.-- s] = compConstr(MtM, Mty,obsLambda, corr, corrLambda, obs, M3d, nmk, encParams, nSnip, M.sub.--s,y.sub.`3 s, weight.sub.-- s, M3dMean)%% compConstr(. . .) -- Computes the matricies with constraints added.%% MtM -- If M.sub.-- s is the encoded matrix of the observation%% data,%% along with the correspondences, then MtM is%% the weighted M.sub.-- s' times the weighted M.sub.-- s.%% Mty -- The weighted M.sub.-- s' times the weighted y.sub.-- s.%% corr -- interactive and tracking correspondences%% corrLambda -- weight value%% obs -- The observation snippet (used to get the head-pos%% data)%% M3d -- Encoded matrix of 3d training data (used for%% constraints)%% nmk -- Number of markers%% encParams -- Parameters used to encode the sequences%% nSnip -- Number of snippets in this sequence%% M.sub.-- s -- List of M's over all snippets%% y.sub.-- s -- List of y.sub.-- s over all snippets%% weight.sub.-- s -- List of weight.sub.-- s over all snippets%%%% This function takes the current MtM matricies and Mty vectors and%%%% weights them by obsLambda, and then adds the correspondence%% constraints (corr), weighted by corrLambda.offset = obs.head.data2d;woff = camToWorld(offset);s = encParams.fps;k = encParams.step;nFr = obs.nframes;if (nargin < 14), M3dMean = M3d(:,1)*0; endif (iscell(MtM) == 0)MtM = {MtM};for i = 2:nSnipMtM{i} = MtM{1};endend%% Weight the observation constraints MtM and Mty by obsLambdaMty = Mty * obsLambda;for snip = 1:nSnipMtM{snip} = MtM{snip} * obsLambda;endfprintf(1, `Adding Marker Constraints . . .`);for c = 1:length(corr)pnew = mean(corr{c}.pos(:,[2 1]), 1);weight = corr{c}.weight;which = corr{c}.mkr;t = corr{c}.t;for snip = 1:nSnipb = (snip-1)*k+1;e = b + s - 1;pnewW = (camToWorld(pnew) - woff(b,:))';% pnewW = (camToworld(pnew)');if ((t >= b) & (t <= e))i = t-b+1;x0 = (which-1)*3 + (i-1)*3*nmk + 1;% disp(sprintf('Marker %d is at (%d, %d) (%f %f, world) in snippet %dat Time %d (with weight %f)', which, round(pnew(1)),round(pnew(2)), pnewW(1), pnewW(3), snip, t, weight));%% Add this constraint to the M.sub.-- s matrixconstr = M3d([x0 x0+2], :) - M3dMean([x0x0+2])*ones(1,size(M3d,2));% constr = M3d([x0 x0+2], :);M.sub.-- s{snip} = {M.sub.-- s{snip}{:}, constr};%% Add the endpoint of the spring (desired value) to the y.sub.-- svectory.sub.-- s{snip} = {y.sub.-- s{snip}{:}, [pnewW(1); pnewW(3)]-M3dMean([x0x0+2])};% y.sub.-- s{snip} = {y.sub.-- s{snip}{:}, [pnewW(1); pnewW(3)]}; %% Add the weight of this correspondence (times the weight %% of all correspondences (corrLambda)) to the weight array weight.sub.-- s{snip} = [weight.sub.-- s{snip} weight*corrLambda]; %% Update MtM and Mty for this snippet MtM{snip} = MtM{snip} + (weight * corrLambda * constr') * constr; Mty(:, snip) = Mty(:, snip) + constr' * (weight * corrLambda * [pnewW(1); pnewW(3)]-M3dMean([x0 x0+2]));endendendfprintf(1, `Done!.backslash.n`);file: /homes/leventon/Matlab/compForm.mfunction F = compForm(I, thetas, smth)% compForm Compute the oriented edge info of an image%% compForm(I, thetas, smth)%% Oct. 9, 1997 wtf corrected C computation.if (nargin < 3), smth = 0.5; end[X, Y, R, T] = grad(smooth(I, smth));Rv = vectorize(R);Rv = Rv(find(Rv>1e1));thresh = median(Rv);% thresh = sqrt(max(max(R)));% thresh = 1000;c = 1;for theta = thetasC = cos(T);S = sin(T);G = 2*(C * cos(theta) + S * sin(theta)). 2;F(:,:,c) = downsample(smooth((C .* (R>thresh)), 0), 0);c = c + 1;endfile: /homes/leventon/Matlab/compFormc.mfunction F = compFormc(I, thetas, sensor.sub.-- pos, sensor.sub.-- oritwosigmasq,cos.sub.-- power)% compFormc Compute the oriented edge info of an image%% compFormc(I, thetas, smth)%% Oct. 9, 1997 wtf corrected C computation.% Oct. 31, 1997 wtf modified from compForm to match the `formc`% processing of the pulse figures. no orientation,% for now% smth = sigma 2 for comparison with twosigmasq% (2*sigma 2).if (nargin < 6), cos.sub.-- power = 2; end%% find gradients[X, Y, R, T] = grad(I);-wtf)reshold (why vectorize here?Rv = vectorize(R);Rv = Rv(find(Rv>1e1));thresh = median(Rv);%% The above can be replaced with thresh = median(R(find(R(:)>1e1)));%% Need to make sure sensor.sub.-- pos fall on integer grid%% Should we interpolate?sensor.sub.-- pos = round(sensor.sub.-- pos);%% Need to make sure in range:which.sub.-- in.sub.-- range = find((sensor.sub.-- pos(1,:) >0) &(sensorpos(2,:) >0) &. . .(sensor.sub.-- pos(1,:) < size(T,2)) & . . .(sensor.sub.-- pos(2,:) < size(T,1)));%% smoothR = smooth((R>thresh), twosigmasq/2);%% sample. add sensor.sub.-- pos to thisc = 1;F = zeros(length(thetas), length(sensor.sub.-- pos));for theta = thetasC = cos(T);S = sin(T);C = (c * cos(theta) + S * sin(theta)). cos.sub.-- power;% (why were the smooth and downsample statements in here in% compForm.m?)C = C .* R;%%% this is real ugly, and only works if sensor.sub.-- pos is on a%%% rectangular grid. How do it the right way??? wtf oct. 31, 1997%%%%%% And why do "2" and "1" subscripts seem to be reversed???%% F should be a column vector of all the responses at each%%sensor%% position, given by sensor.sub.-- pos. Exactly how do that%% indirection?%% Loosly, want F(i,:) = C(sensor.sub.-- pos)F(c,which.sub.-- in.sub.-- range) = C(sub2ind(size(C), . . .sensorpos(2,which.sub.-- in.sub.-- range),sensor.sub.-- pos(1,which.sub.-- in.sub.-- range)));c = c + 1;endfile: /homes/leventon/Matlab/compNorVel.mfunction [V, GH2] = compNorVel(I1, I2, thetas, thresh, GH1)global tgf1global tgf2global tgf3global thf1global thf2global thf3global thf4tgf1 = [0.0094 0.1148 0.3964 -0.0601];tgf1 = [tgf1 -0.9213 tgf1(4:-1:1)];tgf2 = [0.0008 0.0176 0.1660 0.6383];tgf2 = [tgf2 1.0 tgf2(4:-1:1)];tgf3 = [0.0028 0.0480 0.3020 0.5806];tgf3 = [tgf3 0.0 -tgf3(4:-1:1)];thf1 = [0.0098 0.0618 -0.0998 -0.7551];thf1 = [thf1 0.0 -thf1(4:-1:1)];thf2 = [0.0008 0.0176 0.1660 0.6383];thf2 = [thf2 1.0 thf2(4:-1:1)];thf3 = [0.0020 0.0354 0.2225 0.4277];thf3 = [thf3 0.0 thf3(4:-1:1)];thf4 = [0.0048 0.0566 0.1695 -0.1889];thf4 = [thf4 -0.7349 thf4(4:-1:1)];if (nargin < 5)GH1 = [ ];endif (isempty(GH1))G2a1 = conv2(tgfl, tgf2, J1);G2b1 = conv2(tgf3, tgf3, I1);G2c1 = conv2(tgf2, tgf1, I1);H2a1 = conv2(thf1, thf2, I1);H2b1 = conv2(thf4, thf3, I1);H2c1 = conv2(thf3, thf4, I1);H2d1 = conv2(thf2, thf1, I1);elseG2a1 = GH1{1};G2b1 = GH1{2};G2c1 = GH1{3};H2a1 = GH1{4};H2b1 = GH1{5};H2c1 = GH1{6};H2d1 = GH1{7};endG2a2 = conv2(tgf1, tgf2, 12);G2b2 = conv2(tgf3, tgf3, 12);G2c2 = conv2(tgf2, tgf1, 12);H2a2 = conv2(thf1, thf2, 12);H2b2 = conv2(thf4, thf3, 12);H2c2 = conv2(thf3, thf4, 12);H2d2 = conv2(thf2, thf1, 12);[m, n] = size(G2a1);V = zeros(rn, n, length(thetas));thetas = thetas + pi/2;i = 1;for theta = thetasG2t1 = (cos(theta) 2 * G2a1) - (2*cos(theta)*sin(theta) * G2b1) +(sin(theta) 2 * G2c1);H2t1 = (cos(theta) 3 * H2a1) - (3*cos(theta) 2*sin(theta) * H2b1) +(3*cos(theta)*sin(theta) 2 * H2c1) - (sin(theta) 3 * H2d1);G2t2 = (cos(theta) 2 * G2a2) - (2*cos(theta)*sin(theta) * G2b1) +(sin(theta) 2 * G2c2);H2t2 = (cos(theta) 3 * H2a2) - (3*cos(theta) 2*sin(theta) * H2b2) +(3*cos(theta)*sin(theta) 2 * H2c1) - (sin(theta) 3 * H2d2);V(:,:,i) = atan2(H2t1, G2t1) - atan2(H2t2, G2t2);V(:,:,i) = V(:,:,i) + 2*pi*(V(:,:,i) < -pi) - 2*pi*(V(:,:,i) > pi);V(:,:,i) = V(:,:,i) .* ((G2t1. 2 + H2t1. 2) > thresh 2);V(:,:,i) = V(:,:,i.) .* ((G2t2. 2 + H2t2. 2) > thresh 2);1 = 1 +1;endGH2{1} = G2a2;GH2{2} = G2b2;GH2{3} = G2c2;GH2{4} = H2a2;GH2{5} = H2b2;GH2{6} = H2c2,GH2{7} = H2d2;file: /homes/leventon/Matlab/compNorVel2.mfunction [R, T] = compNorVel2(I1, S1, Td1, I2, S2, Td2, thresh)[m, n] = size(S1);Td1 = Td1 .* (S1>thresh);Td2 = Td2 .* (S2>thresh);T = Td1;C1 = cos(Td1);S1 = sin(Td1);C2 = cos(Td2);S2 = sin(Td2);Dp = (C1.*C2 + S1.*S2)R = acos(Dp) * (Dp > 0);dI = I2 - I1;file: /homes/leventon/Matlab/compNorVel4.mfunction [V, GH2] = compNorVel4(I1, I2, thetas, thresh, GH1)global tg4f1global tg4f2global tg4f3global tg4f4global tg4f5global th4f1global th4f2global th4f3global th4f4global th4f5global th4f6tg4f1 = [0.0084 0.0507 0.1084 -0.1231 -0.5729 0.0606];tg4f1 = [tg4f1 0.9344 tg4f1(6:-1:1)];tg4f2 = [0.0001 0.0019 0.0183 0.1054 0.3679 0.7788];tg4f2 = [tg4f2 1.0 tg4f2(6:-:1)];tg4f3 = [0.0028 0.0229 0.0916 0.1186 -0.1839 -0.4867];tg4f3 = [tg4f3 0.0 -tg4f3(6:-1:1)];tg4f4 = [0.0005 0.0060 0.0456 0.1970 0.4583 0.4851];tg4f4 = [tg4f4 0.0 -tg4f4(6:-1:1)];tg4f5 = [0.0012 0.0124 0.0715 0.2059 0.2053 -0.2173];tg4f5 = [tg4f5 -0.5581 tg4f5(6:-1:1)];if (0)th4f1 = [0.0030 -0.0012 -0.0993 -0.2908 0.1006 0.8322];th4f1 = [th4f1 0.0 -th4f1(6:-1:1)];th4f2 = [0.0001 0.0019 0.0183 0.1054 0.3679 0.7788];th4f2 = [th4f2 1.0 th4f2(6:-1:1)];th4f3 = [0.0021 0.0095 -0.0041 -0.1520 -0.3017 0.1161];th4f3 = [th4f3 0.5715 th4f3(6:-1:1)];th4f4 = [0.0004 0.0048 0.0366 0.1581 0.3679 0.3894];th4f4 = [th4f4 0.0 -th4f4(6:-1:1)];th4f5 = [0.0010 0.0077 0.0258 0.0016 -0.1791 -0.3057];th4f5 = [th4fs 0.0 -th4fs(6:-1:1)];th4f6 = [0.0010 0.0108 0.0611 0.1672 0.1237 -0.3223];th4f6 = [th4f6 -0.6638 th4f6(6:-1:1)];endif (nargin < 5)GH1 = [ ];endif (isempty(GH1))G4a1 = conv2(tg4f1, tg4f2, I1);G4b1 = conv2(tg4f3, tg4f4, I1);G4c1 = conv2(tg4f5, tg4f6, I1);G4d1 = conv2(tg4f4, tg4f3, I1);G4e1 = conv2(tg4f2, tg4f1, I1);if (0)H4al = conv2(th4f1, th4f2, I1);H4b1 = conv2(th4f3, th4f4, I1);H4c1 = conv2(th4f5, th4f6, I1);H4d1 = conv2(th4f6, th4f5, I1);H4e1 = conv2(th4f4, th4f3, I1);H4f1 = conv2(th4f2, th4f1, I1);endelseG4a1 = GH1{1};G4b1 = GH1{2};G4c1 = GH1{3};G4d1 = GH1{4};G4e1 = GH1{5};H4a1 = GH1{6};H4b1 = GH1{7};H4c1 = GH1{8};H4d1 = GH1{9};H4e1 = GH1{10};H4f1 = GH1{11};endG4a2 = conv2(tg4f1, tg4f2, I2);G4b2 = conv2(tg4f3, tg4f4, I2);G4c2 = conv2(tg4fs, tg4fs, I2);G4d2 = conv2(tg4f4, tg4f3, I2);G4e2 = conv2(tg4f2, tg4f1, I2);if (0)H4a2 = conv2(th4f1, th4f2, I2);H4b2 = conv2(th4f3, th4f4, I2);H4c2 = conv2(th4f5, th4f6, I2);H4d2 = conv2(th4f6, th4f5, I2);H4e2 = conv2(th4f4, th4f3, I2);H4f2 = conv2(th4f2, th4f1, I2);end[mi, n] = size(I1);V = zeros(m, n, length(thetas));thetas = thetas + pi/2;i = 1;for theta = thetasc = cos(theta);s = sin(theta);kg4 = [c 4; -4*c 3*s; 6*c 2*s 2; -4*c*s 3; s 4];G4t1 -(G4a1*kg4(1))+(G4b1*kg4(2))+(G4c1*kg4(3))+(G4d1*kg4(4))+(G4e1*kg4(5));G4t2 =(G4a2*kg4(1))+(G4b2*kg4(2))+(G4c2*kg4(3))+(G4d2*kg4(4))+(G4e2*kg4(5));if (0)kh4 = [c 5; -5*c 4*s; 10*c 3*s 2; -10*c 2*s 3; 5*c*s 4; -s 5]H4t1 =(H4a1*kh4(1))+(H4b1*kh4(2))+(H4c1*kh4(3))+(H4d1*kh4(4))+(H4e1*kh4(5))+(H4f1*kh4(6));H4t2 =(H4a2*kh4(1))+(H4b2*kh4(2))+(H4c2*kh4(3))+(H4d2*kh4(4))+(H4e2*kh4(5))+(H4f2*kh4(6));elseR = [cos(theta) -sin(theta) ; sin(theta) cos(theta)];si = -3:0.5:3;hmask = zeros(13,13);for hi = 1:13for hj = 1:13p = (R*[si(hi); si(hj)]);hmask(hi, hj) = H4a(p(i), p(2));endendH4t1 = conv2(hmask, I1);H4t2 = conv2(hmask, I2);H4t1 = imCrop(H4t1, size(hmask, 1), size(hmask, 2), size(I1, 1),size(I1, 2));H4t2 = imCrop(H4t2, size(hmask, 1), size(hmask, 2), size(I1, 1),size(I1, 2));G4t1 = imCrop(G4t1, size(hmask, 1), size(hmask, 2), size(I1, 1),size(I1, 2));G4t2 = imCrop(G4t2, size(hmask, 1), size(hmask, 2), size(I1, 1),size(I1, 2));endV(:,:,i) = atan2(H4t1, G4t1) - atan2(H4t2, G4t2);V(:,:,i) = V(:,:,i) + 2*pi*(V(:,:,i) < -pi) - 2*pi*(V(:,:,i) > pi);V(:,:,i) = V(:,:,i) .* ((G4t1. 2 + H4t1. 2) > thresh 2);V(:,:,i) = V(:,:,i) .* ((G4t2. 2 + H4t2. 2) > thresh 2);i = i + 1;endif (0)GH2{1} = G4a2;GH2{2} = G4b2;GH2{3} = G4c2;GH2{4} = G4d2;GH2{5} = G4e2;GH2{6} = H4a2;GH2{7} = H4b2;GH2{8} = H4c2;GH2{9} = H4d2;GH2{10} = H4e2;GH2{11} = H4f2;endfile: /homes/leventon/Matlab/compOriE.mfunction E = compOriE(I1, thetas)tg4f1 = [0.0084 0.0507 0.1084 -0.1231 -0.5729 0.0606];tg4f1 = [tg4f1 0.9344 tg4f1(6:-1:1)];tg4f2 = [0.0001 0.0019 0.0183 0.1054 0.3679 0.7788];tg4f2 = [tg4f2 1.0 tg4f2(6:-1:1)];tg4f3 = [0.0028 0.0229 0.0916 0.1186 -0.1839 -0.4867];tg4f3 = [tg4f3 0.0 -tg4f3(6:-1:1)];tg4f4 = [0.0005 0.0060 0.0456 0.1970 0.4583 0.4851];tg4f4 = [tg4f4 0.0 -tg4f4(6:-1:1)];tg4f5 = [0.0012 0.0124 0.0715 0.2059 0.2053 -0.2173];tg4f5 = [tg4fs -0.5581 tg4fs(6:-1:1)];G4a1 = conv2(tg4f1, tg4f2, I1);G4b1 = conv2(tg4f3, tg4f4, I1);G4c1 = conv2(tg4f5, tg4f5, I1);G4d1 = conv2(tg4f4, tg4f3, I1);G4e1 = conv2(tg4f2, tg4fi, I1);[m, n] = size(11);V = zeros(m, n, length(thetas));thetas = thetas + pi/2;i = 1;for theta = thetasc = cos(theta);s = sin(theta);kg4 = [c 4; -4*c 3*s; 6*c 2*s 2; -4*c*s 3; s 4];G4t1 =G4a1*kg4(1))+(G4b1*kg4(2) )+(G4c1*kg4(3))+(G4d1*kg4(4) )+(G4e1*kg4(5));if (0)kh4 = [c 5; -5*c 4*s; 10*c 3*s 2; -109*c 2*s 3; 5*c*s 4; -s 5];H4t1 =(H4a1*kh4(1))+(H4b1*kh4(2))+(H4c1*kh4(3))+(H4d1*kh4(4))+(H4e1*kh4(5))+(4f1*kh4(6));elseR = [cos(theta) -sin(theta); sin(theta) cos(theta)];si = -3:0.5:3;hmask = zeros(13,13);for hi = 1:13for hj = 1:13p = (R*[si(hi); si(hj)]);hmask(hi, hj) = H4a(p(1), p(2));endendH4t1 = conv2(hmask, I1);H4t1 = imCrop(H4t1, size(hmask, 1), size(hmask, 2), size(I1, 1),size(I1, 2));G4t1 = imCrop(G4t1, size(hmask, 1), size(hmask, 2), size(I1, 1),size(I1, 2));endE(:,:,i) = H4t1. 2 + G4t1. 2;i = i + 1;endfile: /homes/leventon/Matlab/cons.mfunction a = cons(varargin)n = length(varargin);if (n==0)a.struct = [ ];returnendif (mod(n, 2) == 0)a = struct(varargin{:});elsea = varargin{1};f = fieldnames(a);nf = size(f, 1);for i=2:2:nfound = 0;for j = 1:nfif (strcmp(varargin{i}, f(j))) found = 1;endendif (found == 0)a = setfield(a, varargin{i}, varargin{i+1});endendendfile: /homes/leventon/Matlab/corelCoeff.mfunction [xall, xsall] = coreCoeff1(encobs, knl, coeffs, nEv, lambdas,niter, lambdai)nSnip = size(coeffs, 2);%% The training matrix, MM = knl.enc;%% And it's singular value decompositionU = knl.U;V = knl.V;S = sparse(knl.S);%% I only want to use nEv eigenvectors. Zero the remaining eigen%% values.S(nEv+1:size(S,1), nEv+1:size(S,2)) = 0;S = sparse(S);s2 = diag(S.*S);%% For each snippet in the observationfor snip = 1:nSnipCORING.sub.-- SNIPPET = snip% Select out the initial guess of coefficients for this snippet% (computed by SVD)x0 = coeffs(:,snip);% Select out the observation for this snippety = encobs(:,snip);% Compute M'y ( M' = V*S'*U' )% The reason only part of S is used is because less than% the total number of eigenvectors is used.Mty = V*S(1:size(S,2),1:size(S,2)))'U')*y;clear xs;lct = 1;x = x0;for iambda = lambdas% Compute inv(S+lambda*I)SpLi = sparse(size(S,2), size(S,2));SpLi(1:length(s2), 1:length(s2)) = diag(1./(s2 + lambda));SpLi = sparse(SpLi);MtMpL1pLIi = pinv(M'*M + lambda1 + lambda*eye(294));for ct = 1:niter %%% Step 2. Core the coeffs% Parameters of the amount of coring. . . ex = 2; pt = max(x)*3/4;% Computing the coring of the coefficients ax = abs(x); xc = ((ax >= pt) .* x) + (sign(x) .* (ax < pt) .* (ax) . ex/pt (ex-1)); xc = xc .* (xc > 0);% k = 6;% xc = (x < 0) .* (exp(k*(x+log(1/k)/k)) - (1/k)) + (x >= 0) .* %%% Step 3. Minimize %% E = (data - [training set] * coeffs). 2 + %% lambda * (coeffs - cored.sub.-- coeffs). 2 %% See derivation, MEL's notes file "Friday, July 18th, 1997"% x = (V*SpLi*V') * (Mty + lambda*xc);x = MtMpL1pLIi * (Mty + lambda*xc + lambda1);xs{lct, ct} = x;endlct = lct + 1;endxall(:,snip) = x;xsall{snip} = xs;endfile: /homes/leventon/Matlab/coreCoeff.mfunction [xall, xsall] = coreCoeff(encobs, knl, coeffs, nEv, lambdas,niter)nSnip = size(coeffs, 2);%% The training matrix, MM = knl.enc;%% And it's singular value decompositionU = knl.U;V = knl.V;S = sparse(knl.S);%% I only want to use nEv eigenvectors. Zero the remaining eigen%% values.S(nEv+1:size(S,1), nEv+1:size(S,2)) = 0;S = sparse(S);s2 = diag(S.*S);%% For each snippet in the observationor snip = 1:nSnipCORING.sub.-- SNIPPET = snip% Select out the initial guess of coefficients for this snippet% (computed by SVD)x0 = coeffs(:,snip);% Select out the observation for this snippety = encobs(:,snip);% Compute M'y ( M' = V*S'*U' )%% The reason only part of S is used is because less than%% the total number of eigenvectors is used.Mty = V*S(1:size(S,2),1:size(S,2)))'U'*y;clear xs;lct = 1;x = x0;for lambda = lambdas% Compute inv(S+lambda*I)SpLi = sparse(size(S,2), size(S,2)),SpLi(1:length(s2), 1:length(s2)) = diag(1./(52 + lambda));SpLi = sparse(SpLi);for ct = 1:niter%%% Step 2. Core the coeffs% Parameters of the amount of coring . . .ex = 2;pt = max(x)*3/4;% Computing the coring of the coefficientsax = abs(x);xc = ((aX >= pt) .* x) + (sign(x) .* (ax < pt) .*(ax) . ex/pt (ex-1));%%% Step 3. Minimize%% E = (data - [training set] * coeffs). 2 +%% lambda * (coeffs - cored.sub.-- coeffs). 2%% See derivation, MEL's notes file "Friday, July 18th, 1997"x = (V*SpLi*V') * (Mty + lambda*xc);xs{lct, ct}= x;endlct = lct + 1;endxall(:,snip) = x;xsall{snip} = xs;endfile: /homes/leventon/Matlab/data2dToConstr.mfunction ptcorr = data2dToConstr(seq, varargin)data = seqMkrData(seq);[nmkr, two, T] = size(data);if (two .sup..about. = 2)error(`Only works on 2d mkr data`);endframes = 1:T;markers = 1:nmkr;noise = 0;offset = 0;s = 10;k = 5;%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% Process Optionsargc = 1;while (argc <= nargin-1)if (isstr(varargin{argc}) == 0)disp(`Usage Error: Option is not a string`);help seqFiterror(`Usage Error: Option is not a string`);endswitch varargin{argc}case `frames`frames = varargin{argc+1};argc = argc + 2;case `snippetsize`s = varargin{argc+1};argc = argc + 2;case `snippetoffset`k = varargin{argc+1};argc = argc + 2;case `markers`markers = varargin{argc+1};argc = argc + 2;case `noise`noise = varargin{argc+1};argc = argc + 2;case `offset`offset = varargin{argc+1};argc = argc + 2;otherwisedisp(`Usage Error: Bad Option`);help seqReconerror(`Usage Error: Bad Option`);endendptcorr = { };woff = 0*camToWorld(seq.head.data2d);for t = framesfor i = markerspc.mkr = i;pc.t = t;pc.first = t;pc.pos = worldToCam([data(i,1,t) 0 data(i,2,t)] - woff(t,:));pc.pos = pc.pos([2 1]);pc.pos = pc.pos + offset + noise*randn(1,2);pc.cert = 1;pc.hausfrac = 1;pc.weight = 1;ptcorr{length(ptcorr)+1} = pc;endendfile: /homes/leventon/Matlab/decornp.mfunction [u, s, v, E] = decomp2(M, nmk)[u, s, v] = svd(M);[nrow, ncol] = size(M);nfr = nrow/nmk/3;for i = 1:ncolE{i} = [nfr nmk 30; s(i,i)*matricize(u(:,i), 3)];endfile: /homes/leventon/Matlab/decomp2.mfunction [u, s, v, E] = decomp2(M, nmk)[u, s, v] = svd(M);[nrow, ncol] = size(M);nfr = nrow/nmk/2;nEv = min([nrow, ncol]);for i = 1:nEvE{i} = [nfr nmk; s(i,i)*matricize(u(:,i), 2)];endfile: /homes/leventon/Matlab/diffpts.mfunction [dp, mn, mx] = diffpts(A, p1, p2)f = A(1,1);m = A(1,2);hz = A(1,3);[n, lose] = size(A);A = A(2:n,:);n = n - 1;if (n != f * m)error(`Bad AOA`)endframes = 1:f;dp = frames;for i = framesb = (i-1)*m+1;e = i*m;dp(i) = norm(diff(A([p1+b p2+b], :)));endmn = min(dp);mx = max(dp);file: /homes/leventon/Matlab/dori.mfunction [td, s, C2, C3] = dori(I)x = linspace(-2, 2, 10);y = linspace(-2, 2, 10);for i=1:10for j=1:10g2af(i,j) = G2a(x(i), y(j));g2bf(i,j) = G2b(x(i), y(j));g2cf(i,j) = G2c(x(i), y(j));h2af(i,j) = H2a(x(i), y(j));h2bf(i,j) = H2b(x(i), y(j));h2cf(i,j) = H2c(x(i), y(j));h2df(i,j) = H2d(x(i), y(j));endendg2a = conv2(g2af, I);g2b = conv2(g2bf, I);g2c = conv2(g2cf, I);h2a = conv2(h2af, I);h2b = conv2(h2bf, I);h2c = conv2(h2cf, I);h2d = conv2(h2df, I);C2f = 0.5 * (g2af. 2 - g2cf. 2) + 0.46875 * (h2af. 2 - h2df. 2) +0.28125 * (h2bf. 2 - h2cf. 2) +0.1875 * (h2af.*h2cf - h2bf.*h2df);C3f = - g2af.*g2bf - g2bf.*g2cf - 0.9375 * (h2cf.*h2df + h2af.*h2bf) -1.6875 * h2bf .* h2cf - 0.1875 * h2af.*h2df;C2a = conv2(C2f, I);C3a = conv2(C3f, I);C2 = 0.5 * (g2a. 2 - g2c. 2) + 0.46875 * (h2a. 2 - h2d. 2) +0.28125 * (h2b. 2 - h2cf. 2) + 0.1875 * (h2af.*h2fc - h2bf.*h2df);C3f = - g2af.*g2bf - g2bf.*g2cf - 0.9375 * (h2cf.*h2df + h2af.*h2bf) -1.6875 * h2bf.* h2cf - 0.1875 * h2af.*h2df;td = atan2(C2, C3)/2;s = sqrt(C2. 2 + C3. 2);file: /homes/leventon/Matlab/downssmple.mfunction Id = downsample(I, t)if (nargin < 2)t = 1;endif (t < 1)Id = I;returnendif (t > 1)I = downsample(I, t-1);end[m, n, p] = size(I);I = I + I([1:m-1], :, :);I = I(:, [1:n-1], :) + I;I = I + I(2:m m], :, :);I = I(;, [2:n n], :) + I;I = I + I([1 1:m-1], :, :);I = I(:, [1 1:n-1], :) + I;I = I + I(`:m m], :, :);I = I(:, [2:n n], :) + I;% Id = I(1:2:m, 1:2:n, :)/256;Id = I(1:2:m, 1:2:n, :)/128;file: /homes/leventon/Matlab/drawcirc3.mfunction [x, y] = drawcirc(c, r)if (length(c) == 2)elseif (length(c) == 3)u = rand(1,3);v = cross(r, u);if (norm(v) < 0.001)u = rand(1,3);v = cross(r, u);ACK = 1endx = v / norrn(v);u = cross(r, x);y = u / norm(u);x = x * norm(r);y = y * norm(r);% T = 1:10;% for t = T% p = c + x * cos(2*pi*t/10) + y * sin(2*pi*t/10);% X(t) = p(1);% Y(t) = p(2);% Z(t) = p(3);% end% plot3(X, Y, Z);endfile: /homes/leventon/Matlab/drawcyl.mfunction drawcyl(p, o, r)DCYL = [p; o; r];p = diag(diag(p));R = aToR(o);N = 50;T = linspace(0, 2*pi, N);held = ishold;X0 = zeros(N,1);Y0 = zeros(N,1);Z0 = zeros(N,1);X1 = zeros(N,1);Y1 = zeros(N,1);Z1 = zeros(N,1);P0 = [r(1) * cos(T); r(2) * sin(T); zeros(1,N)];P1 = [r(1) * cos(T); r(2) * sin(T); r(3)*ones(1,N)];P0 = R*P0 + (p * ones(1, N));P1 = R*Pi + (p * ones(1, N));% X(1:2:2*N) = P0(1,:);% X(2:2:2*N) = P1(1,:);% Y(1:2:2*N) = P0(2,:);% Y(2:2:2*N) = P1(2,:);% Z(1:2:2*N) = P0(3,:);% Z(2:2:2*N) = P1(3,:);X = [P0(1,:) P1(1,:)];Y = [P0(2,:) P1(2,:)];Z = [P0(3,:) P1(3,:)];X = [P0(1,1:N/2) P1(1,N/2:-1:1) P0(1,N:-1:N/2) P1(1,N/2:N)];Y = [P0(2,1:N/2) P1(2,N/2:-1:1) P0(2,N:-1:N/2) Pi(2,N/2:N)];Z = [P0(3,1:N/2) P1(3,N/2:-1:1) P0(3,N:-1:N/2) P1(3,N/2:N)];plot3(X, Y, Z);if (held == 0)hold offendfile: /homes/leventon/Matlab/drawpart.mfunction drawpart(part, pg, og, deep)if (nargin < 4)deep = 1000000;endDRAWING = part{1};pos1 = zeros(1:6);for i = 5:10if (isempty(part{i}) == 0)pos1(i-4) = part{i};endendp1 = pos1(1:3);o1 = pos1(4:6);Rg = aToR(og);R1 = aToR(ol);Tg = [[Rg; 0 0 0] [diag(diag(pg)); 1]];T1 = [[R1; 0 0 0] [diag(diag(pl)); 1]];T = Tg * T1;o = rToA(T(1:3,1:3));p = T(1:3,4)';r = part{2} part{3} part{4}];drawcyl(p, o, r);hold ondeep = deep -1;if (deep > 0)subparts = part{11};[lose, n] = size(subparts);for i = 1:ndrawpart(subparts{i}, p, o, deep);endendfile: /homes/leventon/Matlab/drawrect.mfunction drawrect (r)%% Draw Rectangle%% r is [p; x; y] where p is center point%% x is x-direction and width of rect%% y is y-direction and length of rectp = r(1, :);r = r(2,:);y = r(3,:);pul = p - x - y;pur = p + x - y;pll = p - x + y;plr = p + x + y;held = ishold;plot([pul(1) pur(1)], [pul(2) pur(2)]);hold onplot([pur(1) plr(1)], [pur(2) plr(2)]);plot([plr(1) pll(1)], [plr(2) pll(2)]);plot([pli(1) pul(1)], [pil(2) pul(2)]),if (held == 0)hold offendfile: /homes/leventon/Matlab/drawrect3.mfunction drawrect3(r)%% Draw 3D Rectangle%% r is [p; x; y; z] where p is center point%% x is x-direction and width of rect%% y is y-direction and length of rect%% z is z-direction and depth of rectp = r(1,:);x = r(2,:);y = r(3,:);z = r(4,:);puli = p - x - y - z;puri = p + x - y - z;plli = p - x + y - z;plri = p + x + y - z;pulo = p - x - y + z;puro = p + x - y + z;pllo = p - x + y + z;plro = p + x + y + z;held = ishold;plot3([puli(1) puri(1)], [puli(2) puri(2)], [puli(3) puri(3)]);hold onplot3([puri(1) plri(1)], [puri(2) plri(2)], [puri(3) plri(3)]);plot3([plri(1) plli(1)], [plri(2) plli(2)], [plri(3) plli(3)]);plot3([plli(1) puli(1)], [plli(2) puli(2)], [plli(3) puli(3)]);plot3([pulo(1) puro(1)], [pulo(2) puro(2)], [pulo(3) puro(3)]);plot3([puro(1) plro(1)], [puro(2) plro(2)], [puro(3) plro(3)]);plot3([plro(1) pllo(1)], [plro(2) pllo(2)], [plro(3) pllo(3)]);plot3([pllo(1) pulo(1)], [pllo(2) pulo(2)], [pllo(3) pulo(3)]);plot3([puli(1) pulo(1)], [puli(2) pulo(2)], [puli(3) pulo(3)]);plot3([puri(1) plro(1)], [puri(2) puro(2)], [puri(3) puro(3)]);plot3([plli(1) pllo(1)], [plli(2) pllo(2)], [plli(3) pllo(3)]);plot3([plri(1) plro(1)], [plri(2) plro(2)], [plri(3) plro(3)]);if (held == 0)hold offendfile: /homes/leventon/Matlab/drawseg3.mfunction H = drawseg3(p1, p2, c)if (nargin > 2)H = plot3([p1(1) p2(1)], [p1(2) p2(2)], [p1(3) p2(3)], c);elseH = plot3([p1(1) p2(1)], [p1(2) p2(2)], [p1(3) p2(3)]);endfile: /homes/leventon/Matlab/encodeVid.mfunction [d, frame] = encodeVid(V, s, k, A)% encodeVid Encode a video sequenceif (iscell(V{1}) == 0)V = {V};endnvid = length(V);d = [ ]; - for w = 1:nvidn = length(V{w}),frame = zeros(30*4, n-1);for t=1:n-1tNV = compNorVel4(V{w}{t:t+1}, [0, pi/4, pi/2, 3*pi/4], 0.1);for thetai = 1:4[nvPyr, nvind] = buildGpyr(NV(:, : ,thetai));vv = vectorize(extractPyr(nvPyr, nvInd, length(nvInd)));length (vv)b = (thetai-1)*30+1;frame(b:b+30-1,t) = vv;endendfor i=1:k:(n-s)ithissnip = [ ];for t=((1:s)+i-1)thissnip = [thissnip; frame(:,t)];endd = [d thissnip];endendfile: /homes/leventon/Matlab/eraseme.mseq = seqCrop(seqGet(6, `motion`, pi/2), 23);enc = seqEncode(seq, knlMo.encParams);nEvInit = 10;snips = 1:3;[rcn, M3d, cc, b, sic] = seqRecon(seq, knlMo, `3d', nEvInit, {10 5 3},knl1d.enc);options(1) = 1; options(2) = 1; options(3) = 1; options(14) = 25;global fEnergy.sub.-- Paramsclear bsbs{1,1} = b;bs{1,2} = b;count = 2;for nEv = [nEvInit 30 50 75 100 125 150 175 200 225 250 275 294]Si = spdiags(1./diag(knlMo.S), 0, size(knlMo.V,1), nEv);S = spdiags(diag(knlMo.S), 0, nEv, size(knlMo.V,1));cc = (S * (knlMo.V'*b));ccc = reshape(cc(:,snips), length(snips)*nEv, 1);M3d = knl3d.enc * knlMo.V * Si;MMo = knlMo.U(:,1:nEv);fEnergyInit(MMo, M3d, enc(:,snips), nEv, .2, 0);[e, oe, ke, lene] = fEnergy(ccc);fEnergy.sub.-- Params.1KE = oe/ke;cccm = fminu(`fEnergy`, ccc, options, `fdEnergy`);b = knlMo.V * (Si * reshape(cccm, nEv, length(snips)));bs{count,1} = b;bs{count,2} = coreCoeff(enc, knlMo, b, nEv, 1000, 1);count = count + 1;endfor i = 1:10for j = [0 pi/2]seq = seqGet(i, `3d`, j);seqi = seqRender(seq);seqm = seqConvert(seqi, `motion`);for k=1:seqi.nframes seqi.data{k} = uint8(seqi.data{k}/4);endsave(sprintf(`Smed/seq-i-%02d-%03d.mat`, i, round(j*180/pi)),`seqi')for k=1:seqm.nframes seqm.data{k} = uint8((seqm.data{k}+pi)/(2*pi)*255);endsave(sprintf(`Smed/seq-m-%02d-%03d.mat`, i, round(j*180/pi)),`seqm`);endend% seqView(seqDecode(knl3d.enc * b, `3d`, {10, 5, 37}));file: /homes/leventon/Matlab/errSmooth.mfunction [e, ea] = errSmooth(f)f = diag(diag(f));n = length(f);M = ones(4,3);M = [1 1 1; 4 2 1; 9 3 1; 16 4 1];ea = zeros(1,n-4);for i = 1:n-4b = f(i:i+3);x = M.backslash.b;ea(i) = ea(i) + normsq(M*x - b);ende = sqrt(mean(ea));file: /homes/leventon/Matlab/errplot.mfunction errplot(x, ym, ys, args)n = length(ym);if (nargin < 4)args = `x-`;endif (isempty(x))x = 1:n;endheld = ishold;ya = ym - ys;ya = max(0, ya);yb = ym + ys;plot(x, ym, args);hold on;for i = 1:nplot([x(i) x(i) x(i)], [ya(i) ym(i) yb(i)], args);endif (held == 0)hold offendfile: /homes/leventon/Matlab/exists.mfunction yn = exists(a, b)f = fields(a);nf = size(f);yn = 0;for i=1:nfif (strcmp(b, f(i)))yn = 1;endendfile: /homes/leventon/Matlab/extractpyr.mfunction I = extractPyr(pyr, ind, lev)b0 = 1;for i = 1:lev-1b0 = b0 + ind(i,1) * ind(1,2);endb1 = b0 + ind(lev,1)*ind(lev,2) - 1;I = matricize(pyr(b0:b1), ind(lev,1))';file: /homes/leventon/Matlab/f1.mfunction v = f1(x)alpha = 1;v = sum(abs(x). alpha);file: /homes/leventon/Matlab/f1dim.m%% f1dim.m p. 419, num rec. in c. Form a 1-d function%% where the input is step along high-dim line, and%% output is multi-dim function value at that point.%% Communicate high-d parameters via global variables.function [f] = f1dim(x);global pcom xicom nrfunc;xt = pcom + x * xicom;f = eval([nrfunc, `(xt)`]);file: /homes/leventon/Matlab/f2.mfunction x = f2(t, R, r, x0)x = abs((R * [r(1)*cos(t); r(2)*sin(t)]) - x0);file: /homes/leventon/Matlab/fCorrE.mfunction e = fcorrE(y)% fprintf(1, `o`);global fGbs.sub.-- Pararnsy = reshape(y, length(y)/fObs.sub.-- Params.nSnip, fObs.sub.-- Params.nSnip);x = fObs.sub.-- Params.VpSpi * y;e = 0;for i=1:fObs.sub.-- Params.nSnipnRep = length(fObs.sub.-- Params.Msnip{i});for j=1:nRepe2 = fObs.sub.-- Params.weightsnip{i}(j) * (fObs.sub.-- Params.obssnip{i}{j} -fObs.sub.-- Params.Msnip{i}{j}* x(:,i));e = e + sum(e2. 2);endende = mean(e);file: /homes/leventon/Matlab/fCorrEInit.mfunction fObsEInit(Msnip, obssnip, weightsnip, MtM, Mty, nEv, VpSpi)global fObs.sub.-- Paramsif (nargin < 7), VpSpi = 1; end% The matrix to multiply x by to get coefficients of training% snippets.fObs.sub.-- Params.VpSpi = VpSpi;% These values are vary from snippet to snippet.fObs.sub.-- Params.Msnip = Msnip;fObs.sub.-- Params.obssnip = obssnip;fObs.sub.-- Params.weightsnip = weightsnip;nSnip = 1;if (iscell(MtM) == 0)MtM = {MtM}Mty = {Mty}for snip = 2:nSnipMtM{snip} = MtM{1};Mty{snip} = Mty{1};endendfObs.sub.-- Params.MtM = MtM;fObs.sub.-- Params.Mty = Mty;fObs.sub.-- Params.nSnip = nSnip;fObs.sub.-- Params.nEv = nEv;file: /homes/leventon/Matlab/fEnergy.mfunction [e, oe, ke, lene] = fEnergy(x, p)global fEnergy.sub.-- Paramsfprintf(1, `<`);oe = fObsE(x);if (fEnergy.sub.-- Params.1KE .sup..about. = 0)ke = fKE(x);elseke = 0;endif (fEnergy.sub.-- Params.lLenE .sup..about. = 0)lene = fLenE(x);elselene = 0;endif (fEnergy.sub.-- Params.1PulseE .sup..about. = 0)pulsee = fPulseE(x);elsepulsee = 0;endfprintf(1, `>`);e = oe + fEnergy.sub.-- Params.1KE*ke + fEnergy.sub.-- Params.1LenE*lene+ . . .fEnergy.sub.-- Params.1PulseE*pulsee;if (nargin == 2)Es = [oe ke lene pulsee]endfile: /homes/leventon/Matlab/fEnergy2.mfunction e = fEnergy(x)%% Usage:%% fEnergyInit(`eName1`), lambda1, {initArgs1}, . . .,%% `eNameN`, lambdaN, {initArgsN})%% When fEnergyInit is called, it, in turn, calls%% f `eName`Init(initArgs1), . . ., f`eNameN(Init(initArgsN)%%%% When fEnergy is called with arg x, it computes the total energy:%% e = lambda1 * f`eName1`(x) + . . . + lambdaN * f`eNameN`(x)%%%% When fdEnergy is called with arg x, it computes the total deriv%% energy:%% de = lambda1 * fd`eName1`(x) + . . . + lambdaN * fd`eNameN`(x)global fEnergy.sub.-- Paramsglobal FOOFOO = FOO + 1;s = ` `;e = 0;for i = 1:fEnergy.sub.-- Params.Nlambda = fEnergy.sub.-- Params.lambdas{i};if (lambda .sup..about. = 0)val = feval([`f` fEnergy.sub.-- Params.names{i}], x);e = e + lambda * val;s = [s sprintf(`%s:%e`, fEnergy.sub.-- Params.names{i}, val)];endends = [s sprintf(`TotE:%e `, e)];if (FOO > 50)FOO = 0;disp(s);endfile: /homes/leventon/Matlab/fEnergyInit.mfunction fEnergyInit(VpSpi, M3d, . . .Mall, obsall, weightall, . . .Msnip, obssnip, weightsnip, . . .MtM, Mty, nEv,. . .pulse.sub.-- u, pulse.sub.-- goal, pulse.sub.-- stacked.sub.-- rep, . ..pulse.sub.-- sensor.sub.-- pos, pulse.sub.-- sensor.sub.-- ori, . . .pulse.sub.-- twosigmasq, pulse.sub.-- cos.sub.-- power, . . .lamKE, lamLenE, lamPulse)global seqINFOglobal fEnergy.sub.-- ParamsC = seqINFO.clengths;%% Initialize the Kinetic Energy functionsfKEinit(M3d);%% Initialize the Length Energy functionsfLenEinit(M3d, C);%% Initialize the (Observation (+ constraints) Energy FunctionsfObsEInit(Mall, obsall, weightall, Msnip, obssnip, weightsnip, MtM,Mty, . . .nEv, VpSpi);%% Initialize the Pulse Energy FunctionsfPulseEInit(pulse.sub.-- u, pulse.sub.-- goal, nEv, pulse.sub.-- stacked.sub.-- rep,pulse.sub.-- sensor.sub.-- pos, . . .pulse.sub.-- sensor.sub.-- ori, pulse.sub.-- twosigmasq, pulse.sub.--cos.sub.-- power);%% Set the lambda weighting parametersfEnergy.sub.-- Params.1KE = lamKE;fEnergy.sub.-- Params.1LenE = lamLenE;fEnergy.sub.-- Params.1PulseE = lamPulse;file: /homes/leventon/Matlab/fEnergyInit2.mfunction fEnergyInit (varargin)%% Usage:%% fEnergyInit(`eName1`, lambda1, {initArgs1}, . . .,%% `eNameN`, lambdaN, {initArgsN})%%%% When fEnergyInit is called, it, in turn, calls%% f`eNamel `Init(initArgs1), . . ., f(`NameN`Init(initArgsN)%%%% When fEnergy is called with arg x, it computes the total energy:%% e = lambda1 * f`eName1`(x) + . . . + lambdaN * fs`eNameN`(x)%%%% When fdEnergy is called with arg x, it computes the total deriv%% energy:%% de = lambda1 * fd`eName1`(x) + . . . + lambdaN * fd`eNameN`(x)global fEnergy.sub.-- Paramsif (mod(nargin, 3) .sup..about. = 0)error(`Usage: fEnergyInit(eName1, lambda1, {initArgs1}, . . ., eNameN,lambdaN, {initArgsN})`);endN = nargin / 3;fEnergy.sub.-- Params.N = N;for i = 1:NfEnergy.sub.-- Params.names{i} = varargin{3*(i-1)+1};fEnergy.sub.-- Params.lambdas{i} = varargin{3*(i-1)+2};feval([`f` fEnergy.sub.-- Params.names{i}`Init`], varargin{3*i}{:});endfile:/homes/leventon/Matlab/fISSDE.mfunction [e, vals, data] = fISSDE(y)global fISSD.sub.-- Params% fISSD.sub.-- Params.count = fISSD.sub.-- Params.count + i;% global ALL.sub.-- YS% ALL.sub.-- YS(:,fISSD.sub.-- Params.count) = y;y = reshape(y, length(y)/fISSD.sub.-- Pararns.nSnip, fISSD.sub.--Params.nSnip);x = fISSD.sub.-- Params.VpSpi * y;e = 0;nmk = fISSD.sub.-- Params.nmk;s = fISSD.sub.-- Params.encParams.fps;% seq = seqDecode(fISSD.sub.-- Params.M * x(:,1) + fISSD.sub.-- Params.MMean, . . .,% `3d`, fISSD.sub.-- Params.encParams, . . .% fISSD.sub.-- Params. seqIm);% data = permute(reshape(seq.data(2:size(seq.data,1), :), nmk, s, 3),% [1 3 2]);M = fISSD.sub.-- Params.M * x(:,1) + fISSD.sub.-- Params.MMean +fISSD.sub.-- Params.offsets;data = permute(reshape(M, 3, nmk, s), [2 1 3]);imsize = [size(fISSD.sub.-- Params.seqIm.data{1}) 1];% Use data([16 17 34 35 36],:,i) for just left arm/handfor i = 1:sX(:,:) = max(1, worldToCam(data(:,:,i)));X(:,1) = min(X(:,1), imsize(2));X(:,2) = min(X(:,2), imsize(1));for j = 1:size(imsize(3))vals(:,i,j) = interp2(fISSD.sub.-- Params.seqIm.data{i}(:,:,j), X(:,1),X(: ,2));endende = mean(mean(mean(diff(vals, 1, 2). 2)));return%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%global fISSD.sub.-- Paramsy = reshape(y, length(y)/fISSD.sub.-- Params.nSnip, fISSD.sub.-- Params.nSnip);x = fISSD.sub.-- Params.VpSpi * y;e = 0;nmk = 37;s = 10;seq = seqDecode(fISSD.sub.-- Params.M * x(:,1) + fISSD.sub.-- Params.MMean, . . .`3d`, fISSD.sub.-- Params.encParams, . . .fISSD.sub.-- Params.seqIm);data = permute(reshape(seq.data(2:size(seq.data,1), :), nmk, s, 3), [13 2]);for i = 1:2X = max(1, round(worldToCam(data(:,:,i))));X(:,1) = min(X(:,i), size(fISSD.sub.-- Params.seqIm.data{i}, 2));X(:,2) = min(X(:,2), size(fISSD.sub.-- Params.seqIm.data{i}, 1));ind(: ,i) = sub2ind(size(fISSD.sub.-- Params.seqIm.data{1}), X(: ,2),X(: ,1));endfor i = 2:2e = e + sum((fISSD.sub.-- Params.seqIm.data{i}(ind(:,i))- . . .fISSD.sub.-- Params.seqIm.data{i-1}(ind(:,i-1))). 2);end% e = mean(e);file: /homes/leventon/Matlab/fISSDEInit.mfunction fISSDEInit(M, VpSpi, seqIm, nSnip, MMean, encParams, smth)global fISSD.sub.-- ParamsfISSD.sub.-- Params.M = M;fISSD.sub.-- Params.nSnip = nSnip;fISSD.sub.-- Params.scale = seqIm.scale;fISSD.sub.-- Params.VpSpi = VpSpi;fISSD.sub.-- Params.MMean = MMean;fISSD.sub.-- Params.head = camToworld(seqIm.head.data2d);fISSD.sub.-- Params.encParams = encParams;fISSD.sub.-- Params.smth = smth;global seqINFOnmk = seqINF0.nummkr(seqIm.subtype);fISSD.sub.-- Params.nmk = nmk;for i = 1:10seqIm.data{i} = smooth(seqIm.data{i}, smth);% seqIm.data{i} = conv2(seqIm.data{i}, [1 4 1; 4 -20 4; 1 4 1],`same`);endfISSD.sub.-- Params.seqIm = seqIm;fISSD.sub.-- Params.offsets =reshape(repmat(camToWorld(seqIm.head.data2d(1, :))`, . . .nmk*encParams.fps, 1), nmk*encParams.fps*3,1);file: /homes/leventon/Matlab/fKE-old.m%% OLD AND OUTDATEDfunction e = fKE(v)fprintf(1, `f`);global fKE.sub.-- ParamsMcon = fKE.sub.-- Params.Mcon;nmk = fKE.sub.-- Params.nmk;s = fKE.sub.-- Params.s;Comb = fKE.sub.-- Params.Comb;nVec = fKE.sub.-- Params.nVec;k = fKE.sub.-- Params.k;nt =- k * nVec + (s-k);n = nt*nmk*3;e = 0;ncoef = size(v, 1)/nVec;or i=1:nVecb = (i-1)*ncoef + 1;Mv(:, i) = Mcon*v(b:b+ncoef-1);endmv = vectorize(Mv`);X = (Comb*mv);for k = 1:nmkt = 2;i = (t-2)*3*nmk + (k-1)*3 + 1;j = (t-1)*3*nmk + (k-1)*3 + 1;V = sparse([1 2 3 1 2 3],[i:i+2 j:j+2],[-1 -1 -1 1 1 1], 3, n);vel1 = V*X;for t = 2:nt-1STATUS = [t k];1 = (t-1)*3*nmk + (k-1)*3 + 1;j = (t)*3*nmk + (k-1)*3 + 1;V = sparse([1 2 3 1 2 3],[i:i+2 j:j+2],[-1 -1 -1 1 1 1], 3, n);vel2 = V*X;dke = (vel2`*vel2 - vel1'*vel1);e = e + (dke 2);vel1 = vel2;endende = e / nmk / (nt-2);file: /homes/leventon/Matlab/fKE.mfunction [e, Xsdd] = fKE(v)%% fKE Compute the Kinetic Energy of the coefficient vector v.fprintf(1, `k`);global fKE.sub.-- Paramsnmk = fKE.sub.-- Params.nmk;s = fKE.sub.-- Params.s;nVec = fKE.sub.-- Params.nVec;k = fKE.sub.-- Params.k;nt = k * nVec + (s-k);ncoef = size(v, 1)/nVec;n = 3*s*nmk*nVec;mv = reshape(fKE.sub.-- Params.Mcon * reshape(v, ncoef, nVec), n, 1);X = (fKE.sub.-- Params.Comb*mv);%% X is an 3*nmk*nt x 1 vectorXs = reshape(X, 3, nmk, nt);Xsd = squeeze(sum(diff(Xs, 1, 3). 2, 1));Xsdd = diff(Xsd, 1, 2). 2;e = mean(Xsdd(:));file: /homes/leventon/Matlab/fKE2.mfunction e = fKE(v)fprintf(1, `f`);global fKE.sub.-- Paramsnmk = fKE.sub.-- Params.nmk;s = fKE.sub.-- Params.s;nVec = fKE.sub.-- Params.nVec;k = fKE.sub.-- Params.k;nt = k * nVec + (s-k);ncoef = size(v, 1)/nVec;n = 3*s*nmk*nVec;mv = reshape(fKE.sub.-- Params.Mcon * reshape(v, ncoef, nVec), n, 1);X = (fKE.sub.-- Params.Comb*mv);%% X is an 3*nmk*nt x 1 vectorXs = reshape(X, 3, nmk, nt);Xsd = squeeze(sum(diff(Xs, 1, 3). 2, 1));Xsdd = diff(Xsd, 1, 2). 2;e = mean(Xsdd(:));file: /homes/leventon/Matlab/fKE.sub.-- old.m%% OLD AND OUTDATEDfunction e = fKE(v)fprint(1 `f`);global fKE.sub.-- ParamsMcon = fKE.sub.-- Params.Mcon;nmk = fKE.sub.-- Params.nmk;s= fKE.sub.-- Params.s;Comb = fKE.sub.-- Params.Comb;nVec = fKE.sub.-- Params.nVec;k = fKE.sub.-- Params.k;nt = k * nVec + (s-k);n = nt*nmk*3;e = 0;ncoef n = size(v, 1)/nVec;for i=1:nVecb = (i-1)*ncoef + 1;Mv(:, i) = Mcon*v(b:b+ncoef-1);endmv = vectorize(Mv');X = (Comb*mv);for k = 1:nmkt = 2;i = (t-2)*3*nmk + (k-1)*3 + 1;j = (t-1)*3*nmk + (k-1)*3 + 1;V = spars([1 2 3 1 2 3],[i:i:+2 j:j+2],[-1 -1 -1 1 1 1], 3, n);vel1 = V*X;for t = 2:nt-1STATUS = [t k];i = (t-1)*3*nml + (k-1(*3 + 1;j = (t)*3*nmk + (k-1)* + 1;V = sparse([1 2 3 1 2 3],]i:i+2 j:j+2],[-1 -1 -1 1 1 1], 3, n);vel2 = V*X;dke = (vel2`*vel2 - vel1`*vel1);e = e + (dke 2);vel1 = bel2;endende = e / nmk / (nt-2);file: /homes/leventon/Matlab/fKEinit.mfunction fKEinit (M)global fKE.sub.-- ParamsfKE.sub.-- Params.Mcon = M;fKE.sub.-- Params.nmk = 37;fKE.sub.-- Params.s = 10;fKE.sub.-- Params.k = 5;fKE.sub.-- Params.nVec = 3;nmk = fKE.sub.-- Params.nmk;s = fKE.sub.-- Params.s;k = fKE.sub.-- Params.k;nVec = fKE.sub.-- Params.nVec;nt = k * nVec + (s-k);snWidth = nt/s;z = linspace(-(s-1)/2, (s-1)/2, s)';sig = s/2-1;mask = exp(-x.*x/(sig*sig));v = zeros(nt, 1);m = mask(1:s/2) ./ (mask(1:s/2)+ mask(s/2+1:s));m2 = vectorize(m* ones(1,nmk*3));N = nmk*3*s;n = nmk*3;md = spdiags([m2(length(m2):-1:1) m2], [0 k*n], k*n, s*n);%% Comb combines nVec snippets into%% e.g.%% s1 is from t=1 . . . 10%% s2 is from t=6 . . . 15%% s3 is from t=16 . . . 20%%%% Then Comb*[s1;s2;s3] == [p1;p2; . . .;p20] (smoothly interpolated)Comb = sparse(snWidth * N, nVec * N);Comb(1:k*n,1:k*n) = speye(k*n);for i=2: (snWidth*2)-1rb = (i-1)*k*n + 1;cb = (i-1)*s*n + 1 - k*n;Comb(rb:rb+k*n-1, cb:cb+s*n-1) = md;endi = snWidth*2;rb = (i-1)*k*n + 1;cb = (i-1)*s*n + 1 - k*n;Comb(rb:rb+k*n-1, cb:cb+k*n-1) = speye(k*n);fKE.sub.-- Params.Comb = Comb;file: /homes/leventon/Matlab/fLenDE.mfunction e = fLenDE(y)global fLenDE.sub.-- Paramsglobal seqINFOy = reshape(y, length(y)/fLenDE.sub.-- Params.nsnip, fLenDE.sub.--Params.nSnip);x = fLenDE.sub.-- Params.VpSpi *y(:,1);nmk = 37;s = 10;data = permute(reshape(fLenDE.sub.-- Params.M * x, 3, nmk, s), [2 1 3]);e = 0;for con=seqINFO.fixed.sub.-- connections);e = e + mean(diff(sum(diff(data(con', :, :)). )). 2);ende = sqrt(e / s);file: /homes/leventon/Matlab/fLenDEInit.mfunction fLenDEInit(M, VpSpi, nSnip)global fLenDE.sub.-- ParamsfLenDE.sub.-- Params.M = M;fLenDE.sub.-- Params.VpSpi = VpSpi;fLenDE.sub.-- Params.nSnip = nSnip;file: /homes/leventon/Matlab/fLenE.mfunction [e, vv] = fLenE(v)fprintf(1, `f`);global fLenE.sub.-- ParamsMcon = fLenE.sub.-- Params.Mcon;C = fLenE.sub.-- Params.C;nmk = fLenE.sub.-- Params.nmk;s = fLenE.sub.-- Params.s;n = nmk*s*3;nC = size(C,1);e = 0;vv = [0 0 0 0];c = 1;for t = 0:s-1b = t*3*37+1;for k = 1:nC1 = 3*(C(k,1)-1) + b;j = 3*(C(k,2)-1) + b;dS = C(k,3);D = sparse([1 2 3 1 2 3],[i:i+2 j:j+2],[-1 -1 -1 1 1 1], 3, n);dif = ((D*Mcon)*v);err = sum(dif.*dif)-dS 2;e = e + (err 2);endende = e / s;file: /homes/leventon/Matlab/fLenEinit.mfunction fLenEinit(M, C)global fLenE.sub.-- ParamsfLenE.sub.-- Params.Mcon = M;fLenE.sub.-- Params.C = C;fLenE.sub.-- Params.nmk = 37;fLenE.sub.-- Params.s = 10;file: /homes/leventon/Matlab/fLenEold.mfunction e = fLenE(v)global fLenE.sub.-- ParamsMcon = fLenE.sub.-- Params.Mcon;C = fLenE.sub.-- Params.connections;nmk = fLenE.sub.-- Params.nmk;s = fLenE.sub.-- Params.s;n = nmk*s*3;nC = size(C,1);e = 0;p = Mcon * v;for t = 0:s-1b = t*37;for i = 1:nCe = e + normsq(p(b+C(i,3)) - (norm(p(b+C(i,1)) - p(b+C(i,2)))));endende = e / s;file: /homes/leventon/Matlab/fObsE.mfunction e = fObsE(y)fprintf(1, `o`);global fObs.sub.-- Paramsy = reshape(y, length(y)/fObs.sub.-- Params.nSnip, fObs.sub.-- Params.nSnip);x = fObs.sub.-- Params.VpSpi * y;e = 0;nRep = length(fObs.sub.-- Params.Mall);for j=1:nRepe = e + sum((fObs.sub.-- Params.weightall(j) * (fObs.sub.-- Params.obsall{j}-fObs.sub.-- Params.Mall{j} * x)). 2);endfor i=1 fObs.sub.-- Params.nSnipnRep = length(fObs.sub.-- Params.Msnip{i});for j=1:nRepe2 = fObs.sub.-- Params.weightsnip{i}(j) * (fObs.sub.-- Params.obssnip{i}{j} -fObs.sub.-- Params.Msnip{i}{j} * x(:,i));e = e + sum(e2. 2);endende = mean(e);file: /homes/leventon/Matlab/fObsEInit.mfunction fObsEInit(Mall, obsall, weightall, Msnip, obssnip,weightsnip, MtM, Mty, nEv, VpSpi)global fObs.sub.-- Paramsif (nargin < 10), VpSpi = 1; end% These values are the same for all snippets.fObs.sub.-- Params.Mall = Mall;fObs.sub.-- Params.obsall = obsall;fObs.sub.-- Params.weightall = weightall;% The matrix to multiply x by to get coefficients of training% snippets.fObs.sub.-- Params.VpSpi = VpSpi;% These values are vary from snippet to snippet.fObs.sub.-- Params.Msnip = Msnip;fObs.sub.-- Params.obssnip = obssnip;fObs.sub.-- Params.weightsnip = weightsnip;nSnip = size(obsall{i},2);if (iscell(MtM) == 0)MtM = {MtM}Mty = {Mty}for snip = 2:nSnipMtM{snip} = MtM{i};Mty{snip} = Mty{i};endendfObs.sub.-- Params.MtM = MtM:fObs.sub.-- Params.Mty = Mty;fObs.sub.-- Params.nSnip = nSnip;fObs.sub.-- Params.nEv = nEv;fObs.sub.-- Params.nElt = size(obsall{1},1) * size(obsall{1},2);file: /homes/leventon/Matlab/fPriorE.mfunction e = fPriorE(alpha3)global fPrior.sub.-- Paramss = 1:length(alpha3);e = alpha3' * fPrior.sub.-- Params.Si(s,s) * alpha3;file: /homes/leventon/Matlab/fPriorEInit.mfunction fPriorEInit(Si)global fPrior.sub.-- ParamsfPrior.sub.-- Params.Si = Si;file: /homes/leventon/Matlab/fPulseE.mfunction out = fPulseE(bet)%% fPulseE (formerly f.sub.-- body.sub.-- bet.m)%% function to be minimized for body problem gradient%% descent to find the body pose that comes closest to%% matching some desired sensor response.%% bet are the coefficients of the linear parmeterization.%% Like f.sub.-- line.m, except a function of coefficient, not x.%% history: oct. 15 1997 freeman created from f.sub.-- line.sub.-- bet.m%% oct. 17, 1997 wtf used formc and derivs%% representations.%% nov. 2, 1997 mel changed this function to fPulseE to%% fit in the current energy framework%% u(:,1:length(bet)) * bet is [pt.sub.-- pos; pt.sub.-- ori] all stackedinto a%% vector.% bet the linear weight coefficients for sequence.%% seq.sub.-- goal the desired response sequence.% u matrix goes from linear weights, bet, to stacked pulse% rep% stacked.sub.-- rep is either `stackedpulse no.sub.-- normals`, or% `stackedpulse`, depending on whether the normals are% included in the stacking.% fprintf(1, `p`);global fPulseE.sub.-- Paramsseq3d = seqDecode(fPulseE.sub.-- Params.u(:,1:length(bet)) * bet + . . .fPulseE.sub.-- Params.M3dMean, . . .fPulseE.sub.-- Params.encParams,fPulseE.sub.-- Params.seq.sub.-- goal);fPulseE.sub.-- Params.seq3d = seq3d;% convert from stackedpulse to sensor responses.fPulseE.sub.-- Params.seq.sub.-- response = . . .seqConvert(seq3d, . . .`formc`, fPulseE.sub.-- Params.sensor.sub.-- pos,fPulseE.sub.-- Params.sensor.sub.-- ori, . . .fPulseE.sub.-- Params.twosigmasq, fPulseE.sub.-- Params.cos.sub.--power);% get total squared error.out = 0;MY.sub.-- EPS = 1e-8;for frame=1:fPulseE.sub.-- Params.seq.sub.-- response.nframesresp = fPulseE.sub.-- Params.seq.sub.-- response.data{frame,1};goal = fPulseE.sub.-- Params.seq.sub.-- goal.data{frame,1};% goal = goal / sum(goal) * sum(resp);out = out + sum(sum((resp - goal). 2));endglobal RESPglobal GOALRESP = [fPulseE.sub.-- Params.seq.sub.-- response.data{:,1};];GOAL = [fPulseE.sub.-- Params.seq.sub.-- goal.data{:,1};];fPulseE.sub.-- Params.count = fPulseE.sub.-- Params.count + 1;if (mod(fPulseE.sub.-- Params.count, 100) == 0)clffor frame=1:fPulseE.sub.-- Params.seq.sub.-- response.nframessubplot(5,2,frame)a = fPulseE.sub.-- Params.seq.sub.-- response.data{frame,1}(:); b = fPulseE.sub.-- Params.seq.sub.-- goal.data{frame,1}(:); hold off plot (a/max (a))% plot(a) hold on plot(b/max(b), `r`)% plot(b, `r`) title(sprintf(`%0.3f/%0.3f`, max(a),max(b)))endpause(0.1);endreturn;file: /homes/leventon/Matlab/fpulseEInit.mfunction fPulseEInit(kn13d, seq.sub.-- goal, nEv, stacked.sub.-- rep, . ..sensor.sub.-- pos, sensor.sub.-- ori, twosigmasq, cos.sub.-- power)global fPulseE.sub.-- Params% Si = spdiags(1./diag(knlpulse.S), 0, size(knlPulse.V,1), nEv);% fPulseE.sub.-- Params.u = knlPulse.enc{1} * knlPulse.V * Si;fPulseE.sub.-- Params.u = knl3d.U;fPulseE.sub.-- Params.seq.sub.-- goal = seq.sub.-- goal;fPulseE.sub.-- Params.stacked.sub.-- rep = stacked.sub.-- rep;fPulseE.sub.-- Params.sensor.sub.-- pos = sensor.sub.-- pos;fPulseE.sub.-- Params.sensor.sub.-- ori = sensor.sub.-- ori;fPulseE.sub.-- Params.twosigmasq = twosigmasq;fPulseE.sub.-- Params.cos.sub.-- power = cos.sub.-- power;fPulseE.sub.-- Params.count = 0;fPulseE.sub.-- Params.encparams = knl3d.encParams;fPulseE.sub.-- Params.M3dMean = knl3d.mean{1};file: /homes/leventon/Matlab/fPulseKalE.mfunction out = fPulseE(bet)%% fPulseE (formerly f.sub.-- body.sub.-- bet.m)%% function to be minimized for body problem gradient%% descent to find the body pose that coines closest to%% matching some desired sensor response.%% bet are the coefficients of the linear parmeterization.%% Like f.sub.-- line.m, except a function of coefficient, not x.%% history: oct. 15 1997 freeman created from f.sub.-- line.sub.-- bet.m%% oct. 17, 1997 wtf used formc and derivs%% representations.%% nov. 2, 1997 mel changed this function to fPulseE to%% fit in the current energy framework%% u(:,1:length(bet)) * bet is [pt.sub.-- pos; pt.sub.-- ori all stackedinto a%% vector.% bet the linear weight coefficients for sequence.% seq.sub.-- goal the desired response sequence.% u matrix goes from linear weights, bet, to stacked pulse% rep% stacked.sub.-- rep is either `stackedpulse.sub.-- no.sub.-- normals`, or `stackedpulse`, depending on whether the normals are included in the stacking.fprintf(1, `p`);global fPulseE.sub.-- Params% now convert to a stackedpulse sequence for which we can find the% response, and compare with the goal response.fPulseE.sub.-- Params.seq.sub.-- response = fPulseE.sub.-- Params.seq.sub.-- goal;fPulseE.sub.-- Params.seq.sub.-- response.type = fPulseE.sub.-- Params.stacked.sub.-- rep;fPulseE.sub.-- Params.seq.sub.-- response.data = fPulseE.sub.-- Params.u(:,1:length(bet)) *bet;% convert frorn stackedpulse to sensor responses.fPulseE.sub.-- Params.seq response = seqConvert(fPulseE.sub.-- Params.seq.sub.-- response,. . .`formc`, fPulseE.sub.-- Params.sensor.sub.-- pos, fPulseE.sub.-- Params.sensor.sub.-- ori,. . .fPulseE.sub.-- Params.twosigmasq, fPulseE.sub.-- Params.cos.sub.--power);% get total squared error.out = 0;for frame=1:fPulseE.sub.-- Params.seq.sub.-- response.nframesout = out + . . .1/obs.sub.-- noise * . . .sum(sum((fPulseE.sub.-- Params.seq.sub.-- response.data{frame,1} - . . .fPulseE.sub.-- Params.seq.sub.-- goal.data{frame,1}). 2));endout = out + (bet' * fPulseE.sub.-- Params.Spi(1:length(bet),1:length(bet)) *bet);return;file: /homes/leventon/Matlab/fRenModelE.mfunction 8 err, v] = fRenModelE(b)global fRenModelE.sub.-- Datadnsamp = fRenModelE.sub.-- Data.dnsamp;smth = fRenModelE.sub.-- Data.smth;fprintf(1, `[`);dat = stickify(fRenModelE.sub.-- Data.M3d*b);for i = 1:10fprintf(1, `.`);dat{i} = smooth(downsample(dat{i}, dnsamp), smth);v(:,:,;,i) = transpose2(dat{i});enderr = sum((v(:)-fRenModelE.sub.-- Data.formv). 2);fprintf(1, `]`);file: /homes/leventon/Matlab/fRenModelEInit.mfunction fRenModelEInit(M3d, seqFo, dnsamp, smth)global fRenModelE.sub.-- DatafRenModelE.sub.-- Data.M3d = M3d;fRenModelE.sub.-- Data.dnsamp = dnsamp;fRenModelE.sub.-- Data.smth = smth;fRenModelE.sub.-- Data.formv = seqEncode(seqFo, `downsample`, dnsamp,`force`, 1, `smooth`, smth);file: /homes/leventon/Matiab/fRenModelEcm.mfunction [err, v] = fRenModelEcm(b)global fRenModelE.sub.-- Datadnsamp = fRenModelE.sub.-- Data.dnsamp;smth = fRenModelE.sub.-- Data.smth;fprintf(1, `[`);pts = fRenModelE.sub.-- Data.M3d*b;pts = pts - reshape(pts(10:12) * ones(1, 370), 1110, 1);dat = stickify(pts);for i = 1:10fprintf(1, `.`);dat{i} = smooth(downsample(dat{i}, dnsamp), smth);v(:,:,:,i) = transpose2(dat{i});enderr = sum((v(:)-fRenModelE.sub.-- Data.formv). 2);fprintf(i, `]`);file: /homes/leventon/Matlab/fRenModelEcmEig.mfunction [err, v] = fRenModelEcnEig(c)global fRenModelE.sub.-- Datadnsamp = fRenModelE.sub.-- Data.dnsamp;smth = fRenModelE.sub.-- Data.smth;nEv = fRenModelE.sub.-- Data.nEv;fprintf(1, `[`);pts = fRenModelE.sub.-- Data.M*c;pts = pts - reshape(pts(10:12) * ones(1, 370), 1110, 1);dat = stickify(pts, dnsamp);for i = 1:10fprintf(1, `.`);dat{i} = smooth(dat{i}, smth);v(:,:,:,i) = (dat{i});enderr = sum((v(:)-fRenModelE.sub.-- Data.formv). 2);fprintf(1, `]`);file: /homes/leventon/Matlab/fRenModelEcmEigInit.mfunction fRenModelEcmInit(M3d, knl, nEv, seqFo, dnsamp, smth)global fRenModelE.sub.-- DataSi = spdiags(1./diag(knl.S), 0, size(knl.V,1), nEv);fRenModelE.sub.-- Data.M = M3d * knl.V * Si;fRenModelE.sub.-- Data.nEv = nEv;fRenModelE.sub.-- Data.dnsamp = dnsamp;fRenModelE.sub.-- Data.smth = smth;fRenModelE.sub.-- Data.formv = seqEncode(seqFo, `downsample`, dnsamp,`force`, 1, `smooth`, smth);file: /homes/leventonlMatlab/fRenModelEcmInit.mfunction fRenModelEcmInit(M3d, seqFo, dnsamp, smth)global fRenModelE.sub.-- DatafRenModelE.sub.-- Data.M3d = M3d;fRenModelE.sub.-- Data.dnsamp = dnsamp;fRenModelE.sub.-- Data.smth = smth;fRenModelE.sub.-- Data.formv = seqEncode(seqFo, `downsample`, dnsamp,`force`, 1, `smooth`, smth);file: /homes/leventon/Matlab/fdCorrE.mfunction de = fdObsE(y)% fprintf(1, `o`);global fObs.sub.-- Paramsy = reshape(y, length(y)/fObs.sub.-- Params.nSnip, fObs.sub.-- Params.nSnip);x = fObs.sub.-- Params.VpSpi * y;de = zeros(size(x));for i=1:fObs.sub.-- Params.nSnipnRep = length(fObs.sub.-- Params.Msnip{i});for j=1:nRepif (isempty(fObs.sub.-- Params.Msnip{i}{j}) == 0)de2 = fObs.sub.-- Params.weightsnip{i}(j) * . . .(fObs.sub.-- Params.obssnip{i}{j} - fObs.sub.-- Params.Msnip{i}{j} *x(: ,i));de2 = fObs.sub.-- Params.Msnip{i}{j}' * de2;de = de + sum(de2,2);endendendde = -2 * fObs.sub.-- Params.VpSpi' * mean(de,2);returnglobal fObs.sub.-- Paramsfprintf(1, `0`);y = reshape(y, length(y)/fObs.sub.-- Params.nSnip, fObs.sub.-- Params.nSnip);x = fObs.sub.-- Params.VpSpi * y;e = zeros(size(y));for i = 1:fObs.sub.-- Params.nSnip` * . . .e(:,i) = e(:,i) + fObs.sub.-- Params.VpSpi' * . . .(fObs.sub.-- Params.Mty(:,i) - fObs.sub.-- Params.MtM{i} * x(:,i));ende = -2 * e(:);file: /homes/leventon/Matlab/fdEnergy.mfunction de = fdEnergy(x)global fEnergy.sub.-- Paramsfprintf(1, `[`);doe = fdObsE(x);if (fEnergy.sub.-- Params.1KE .sup..about. = 0)dke = fdKE(x);elsedke = 0;endif (fEnergy.sub.-- Params.1LenE .sup..about. = 0)dlene = fdLenE(x);elsedlene = 0;endif (fEnergy.sub.-- Params.1PulseE .sup..about. = 0)dpulsee = fdPulseE(x);elsedpulsee = 0;endfprintf(1, `].backslash.n`);de = doe + fEnergy.sub.-- Params.1KE*dke + fEnergy.sub.-- Params.1LenE*dlene + . . .fEnergy.sub.-- Params.lPulseE*dpulsee;file: /homes/leventon/Matlab/fdEnergy2.mfunction de = fdEnergy(x)%% Usage:%% fEnergyInit(`eName1`, lambda1, {initArgs1}, . . .,%% `eNameN`, lambdaN, {initArgsN})%%%% When fEnergyInit is called, it, in turn, calls%% f`eName1`Init(initArgs1), . . ., f`eNameN`Init(initArgsN)%%%% When fEnergy is called with arg x, it computes the total energy:%% e = lambda1 * f`eName1`(x) + . . . + lambdaN * f`eNameN`(x)%%%% When fdEnergy is called with arg x, it computes the total deriv%% energy:%% de = lambda1 * fd`eName1`(x) + . . . + lambdaN * fd`eNameN`(x)global fEnergy.sub.-- Paramsde = 0;for i = 1:fEnergy.sub.-- Params.Nlambda = fEnergy.sub.-- Params.lambdas{i};if (lambda .sup..about. = 0)this.sub.-- de = lambda * feval([`fd` fEnergy.sub.-- Params.names{i}],x);de = de + this.sub.-- de;endendfile: /homes/leventon/Matlab/fdKE.mfunction [e, MMCombt] = fdKE(v)fprintf(i, `k`);global fKE.sub.-- ParamsMcon = fKE.sub.-- Params.Mcon;nmk = fKE.sub.-- Params.nmk;s = fKE.sub.-- Params.s;Comb = fKE.sub.-- Params.Comb;nVec = fKE.sub.-- Params.nVec;k = fKE.sub.-- Params.k;nt = k * nVec + (s-k);n = nt*nmk*3;e = 0;ncoef = size(v)/nVec;mv = reshape(fKE.sub.-- Params.Mcon * reshape(v, ncoef, nVec),3*s*nmk*nVec,1);X = (Comb*inv);e = 0;for k = 1:nmkt = 2;1 = (t-2)*3*nmk + (k-1)*3 + 1;j = (t-1)*3*nmk + (k-1)*3 + 1;V1 = sparse([1 2 3 1 2 3],[i:i+2 j:j+2],[-1 -1 -1 1 1 1], 3, n);vel1 = V1*X;for t = 2:nt-1STATUS = [t k];i = (t-1)*3*nmk + (k-1)*3 + 1;j = (t)*3*nmk + (k-1)*3 + 1;V2 = sparse([1 2 3 1 2 3],[i:i+2 j:j+2],[-1 1 -1 1 1 1], 3, n);vel2 = V2*X;val = (vel2`*vel2 - vel1`*vel1);dke = val * (((V2'*V2)-(V1'*V1)) * X);e = e + dke;vel1 = vel2;V1 = V2;endendval = reshape(Mcon'* reshape(Comb'*e, 1110, 3), 294*3, 1);e = 4 * val / nmk / (nt-2);file: /homes/leventon/Matlab/fdK(E2.mfunction [e, MMCombt] = fdKE(v)fprintf(1, `d`);global fKE.sub.-- ParamsMcon = fKE.sub.-- Params.Mcon;nmk = fKE.sub.-- Params.nmk;s = fKE.sub.-- Params.s;Comb = fKE.sub.-- Params.Comb;nVec = fKE.sub.-- Params.nVec;k = fKE.sub.-- Params.k;nt = k * nVec + (s-k);n = nt*nmk*3;e = 0;fprintf(1, `.`);ncoef = size(v)/nVec;for i=1:nVecb = (i-1)*ncoef + 1;Mv(:, i) = Mcon*v(b:b+ncoef-1);endmv = vectorize(Mv');fprintf(1, `<`);X = (Comb*mv);%% X is an 3*nmk*nt x 1 vectorXs = reshape(X, 3, nmk, nt);Xsd = squeeze(sum(diff(Xs, 1, 3). 2, 1));sqrtval = diff(Xsd, 1, 2)Xsr = reshape(X, 3*nmk, nt);XsrM1 = [Xsr(:, i) Xsr(:, i:nt)]XsrP1 = [Xsr(:, i:nt) Xsr(:, nt)](Xsr - XsrM1). 2 + (XsrP1 - Xsr). 2 - (XsrP1 - Xvect =e = 4 * sqrtval * vect / nmk / (nt-2);n = nt*nmk*3;e = mean(Xsdd(:));fprintf(1,`>`);for k = 1:nmkfprintf(1,`!`);t = 2;1 = (t-2)*3*nmk + (k-1)*3 + 1;j = (t-1)*3*nmk + (k-1)*3 + 1;V1 = sparse([1 2 3 1 2 3],[i:i+2 j:j+2],[-1 -1 -1 1 1 1], 3, n);vel1 = V1*X;for t = 2:nt-1STATUS = [t k];i = (t-1)*3*nmk + (k-1)*3 + 1;j = (t)*3*nmk + (k-1)*3 + i;V2 = sparse([1 2 3 1 2 3],[i:i+2 j:j+2],[-1 -1 -1 1 1 1], 3, n);vel2 = V2*X;val = (vel2'*vel2 - vel1'*vel1);dke = val * (((V2'*V2)-(V1'*V1)) * X);e = e + dke;vel1 = vel2;V1 = V2;endendval = reshape(Mcon'* reshape(Comb'*e, 1110, 3), 294*3, 1);e = 4 * val / nmk / (nt-2);file: /homes/leventon/Matlab/fdLenE.mfunction de = fdLenE(v)fprintf(i, `d`);global fLenE.sub.-- ParamsMcon = fLenE.sub.-- Params.Mcon;C = fLenE.sub.-- Params.C;nmk = fLenE.sub.-- Params.nmk;s = fLenE.sub.-- Params.s;n = nmk*s*3;nC = size(C,1);de = zeros(size(v));for t = 0:s-1b = t*3*37+1;for k = 1:nCi = 3*(C(k,1)-1) +b;j = 3*(C(k,2)-1) + b;dS = C(k,3);D = sparse([1 2 3 1 2 3], [i:i+2 j :j+2],[-1 -1 -1 1 1 1], 3, n);DM = (D*Mcon);dif = (DM*v);de = de + (dif'*dif - dS 2)*(DM'*dif);endendde = 4 * de / s;file: /homes/leventon/Matlab/fdObsE.mfunction e = fdObsE(y)global fObs.sub.-- Paramsfprintf(1, `o`);y = reshape(y, length(y)/fObs.sub.-- Params.nSnip, fObs.sub.-- Params.nSnip);x = fObs.sub.-- Params.VpSpi * y;e = zeros(size(y));for i = 1:fObs.sub.-- Params.nSnipe(:,i) = e(:,i) + fObs.sub.-- Params.VpSpi' * . . .(fObs.sub.-- Params.Mty(:,i) - fObs.sub.-- Params.MtM{i} * x(:,i));ende = -2 * e(:);file: /homes/leventon/Matlab/fdPriorE.mfunction de = fdPriorE(alpha3)global fPrior.sub.-- Paramss = 1:length(alpha3);de = 2 * fprior.sub.-- Params.Si(s,s) * alpha3;file: /homes/leventon/Matlab/fsPulseE.mfunction out.sub.-- bet.sub.-- crop = fdpulseE(bet)%% fdPulseE (formerly grad.sub.-- body.sub.-- bet.rn)%% gradient function to be minimized for descent work%% with body linear combination representation.%% Like grad.sub.-- line bet.m, except as a function of coefficient bet.%% history: oct. 15, 1997 wtf created from grad.sub.-- line.sub.-- bet.%% oct. 17, 1997 wtf used formc and derivs%% representations.%% note: this prograrn assumes that f.sub.-- body.sub.-- bet has been%% called with the same bet value just before this. To%% provide seq.sub.-- response from common.%% otherwise, do: seq.sub.-- response = seqConvert(seq.sub.-- derivs,%% `formc`);%% Oct. 31, 1997 wtf added sensor.sub.-- pos, etc variables in%% command line, and removed the call to tmpsetup to read%% those in.% stacked.sub.-- rep is either `stackedpulse.sub.-- no.sub.-- normals`, or% `stackedpulse`, depending on whether the normals are% included in the stacking.global fPulseE.sub.-- Paramsfprintf(1, `p`);% now convert to a stackedpulse sequence for which we can find the% response, and compare with the goal response.seq.sub.-- derivs = fPulseE.sub.-- Params.seq.sub.-- goal;seq.sub.-- derivs.type = fPulseE.sub.-- Params.stacked.sub.-- rep;seq.sub.-- derivs.data = fPulseE.sub.-- Params.u(:,1:length(bet)) * bet;%% to make sure we have the correct seq.sub.-- response.%% But it doesn't seem necessary; performance is faster in the fminu%% search if I don't repeat this call.%% PulseE.sub.-- Params.seq.sub.-- response = seqConvert(seq.sub.--derivs, `formc`, . . .%% fPulseE.sub.-- Params.sensor.sub.-- pos, fPulseE.sub.-- Params.sensor.sub.-- ori, . . .% fPulseE.sub.-- Params.twosigmasq, fPulseE.sub.-- Params.cos.sub.--power);% convert from stackedpulse to sensor derivatives.seq.sub.-- derivs = seqConvert(seq.sub.-- derivs, `derivs`, . . . fPulseE.sub.-- Params.sensor.sub.-- pos, fPulseE.sub.-- Params.se nsor.sub.-- ori, . . .fPulseE.sub.-- Params.twosigmasq, fPulseE.sub.-- Params.cos.sub.--power);%% the deriv of out = sum(sum((y - y.sub.-- goal). 2));[a,b] = size(fPulseE.sub.-- Params.seq.sub.-- response.data{1,1});[c, tmp] = size(seq.sub.-- derivs.data{1,1});out = [ ];for frame=1:seq.sub.-- derivs.nframesresp = fPulseE.sub.-- Params.seq.sub.-- response.data{frame,1};goal = fPulseE.sub.-- Params.seq.sub.-- goal.data{frame, 1};% goal = goal / max(goal) * max(resp);out = [out; . . .sum(2 * repmat( reshape(resp - goal, 1, a*b), c, 1) .* . . . seq.sub.-- derivs.data{frame, 1},2)];endout.sub.-- bet = fPulseE.sub.-- Params.u' * out;out.sub.-- bet.sub.-- crop = out.sub.-- bet(1:length(bet),:);file: /homes/leventon/Matlab/fdpulseKalE .mfunction out.sub.-- bet crop = fdpulseE(bet)%% fdPulseE (formerly grad.sub.-- body.sub.-- bet.m)%% gradient function to be minimized for descent work%% with body linear combination representation.%% Like grad.sub.-- line bet.m, except as a function of coefficient bet.%% history: oct. 15, 1997 wtf created from grad.sub.-- line.sub.-- bet.%% oct. 17, 1997 wtf used formc and derivs%% representations.%% note: this program assumes that f.sub.-- body.sub.-- bet has been%% called with the same bet value just before this. To%% provide seq.sub.-- response from common.%% otherwise, do: seq.sub.-- response = seqConvert(seq.sub.-- derivs,%% `form`);%% Oct. 31, 1997 wtf added sensor.sub.-- pos, etc variables in%% command line, and removed the call to tmpsetup to read%% those in.% stacked.sub.-- rep is either `stackedpulse.sub.-- no.sub.-- normals`, or% `stackedpulse`, depending on whether the normals are% included in the stacking.global fPulseE.sub.-- Paramsfprint(1, `p`);% now convert to a stackedpulse sequence for which we can find the% response, and compare wiht the goal response.seq.sub.-- derivs = fPulseE.sub.-- Params,seq.sub.-- goal;seq.sub.-- derivs.type = fPulseE.sub.-- Params.stacked.sub.-- rep;seq.sub.-- derivs.data = fOulse.sub.-- Params.u(:,1:length(bet) * bet;%% to make sure we have the corect seq.sub.-- response.%% But it doesn't seem necessary; performance is faster in the fminu%% search if I don't repeat this call. - %% fPulseE.sub.-- Params.seq.sub.-- response = seqConvert(seq.sub.-- drivs, `formc`);/ - % convert fromstackedpulse to sensor derivatives. - seq.sub.-- derivs= seqConvert(seq.sub.-- derivs, `derivs`, . . . fPulseE.sub.-- Params.sensor.sub.-- pos, fPulseE.sub.-- Params.se nsor.sub.-- ori, . . .fPulseE.sub.-- Params.twosigmasq, fPulseE.sub.-- Params.cos.sub.--power);%% the deriv of out = sum(sum((y - y.sub.-- goal), 2));[a,b] = size(fPulseE.sub.-- Params.seq.sub.-- response.data{1,1};[c, tmp] = size(seq.sub.-- derivs.data{1,1} );out = [ ];for frame=1:seq.sub.-- derivs.nframesout = [out; . . .sum(2 * repmat( reshape(fPulseE.sub.-- Params.seq.sub.-- response.data{frame,1} -. . .fPulseE.sub.-- Params.seq.sub.-- goal.data{frame,1} . . . , 1, a*b), c, 1) .* seq.sub.-- derivs.data{frame,1},2)];endout.sub.-- bet = fPulseE.sub.-- Params.u' * out;out.sub.-- bet.sub.-- crop = out.sub.-- bet(1:length(bet),:);out.sub.-- bet.sub.-- crop = out.sub.-- bet.sub.-- crop + . . .fPulseE.sub.-- Params.Spi(1:length(bet), 1:length(bet)) * bet;file: /homes/leventon/Matlab/field.mfunction field(R, T, h)R=R';T=T';[m, n] = size(R);R = R / max(max(R));D(:,:,1) = -R.*sin(T);D(:,:,2) = -R.*cos(T);if (nargin < 3)tl,3 hold offendplot(0, 0);hold onfor i = 1:mfor j = 1:nplot([i i+D(i, j, 1)], -[j j+D(i, j, 2)], `b`)plot(i, -j, `r.`);endendaxis imagefile: /homes/leventon/Matlab/foo.mfunction fooglobal helloload hello.mathellofile: /homes/leventon/Matlab/frameaoa.mfunction frameaoa(A, i, cl)%%if (nargin < 3)cl = [ ];endf = A(1,1);m = A(1,2);hz = A(1,3);[n, lose] = size(A);A = A(2:n,:);n = n - 1;if (n != f * m)error(`Bad AOA`)endxmin = min(A(:,1));xmax = max(A(:,1));ymin = min(A(:,2));ymax = max(A(:,2));zmin = min(A(:,3));zmax = max(A(:,3));hold offb = (i-1)*m+1;e = i*m;plots(A(b:e,1), A(b:e,2), A(b:e,3), `o`);axis([xmin xmax ymin ymax zmin zmax]);% view(0,90);xlabel(`x`); ylabel(`y`); zlabel(`z`);hold onn = size(cl);for i = 1:n[c, v, r] = makecyl(A(b+cl(i,1), :), A(b+cl(i,2), :), cl(i,3));drawcyl(c, v, r);endpause(0. 1);file: /homes/leventon/Matlab/fred.mseqInitload /homes/leventon/Data/fred.matSEQUENCES = 1:10;VIEW.sub.-- ANGLES = (0:3)*pi/2;sel = seqSelect(SEQUENCES VIEW.sub.-- ANGLES);knlFo = seqTrain(sel, `form`, `submean`, 1, `MtM`, 0, `headalign`, 1);knlPulse = seqTrain(sel, `pulse`, `submean`, 1, `MtM`, 0, `headalign`,1);barisli =seqRecon (file: /homes/leventon/Matlab/frprmn.m%% frprrmn.m fletcher-reeves-polak-ribiere minimization. p. 423, num%% in C. Conjugate gradient descent.function [p, fret, iter] = frprmn(p, ftol, func, dfunc);% ITMAX = 200;ITMAX = 1;EPS = 1.0e-10;fp = eval([func, `(p)`]);xi = eval([dfunc, `(p)`]);g = -xi;h = g; xi = g;for iter = 1:ITMAX[p, xi, fret] = linmin(p, xi, func);if (2.0*abs(fret-fp) <= ftol * (abs(fret)+abs(fp) + EPS))return;end;fp = eval([func, `(p)`]);xi = eval([dfunc, `(p)`]);dgg = 0.0; gg = 0.0;gg = sum(g. 2);%% for fletcher-reeves:%% dgg = sum(xi. 2);%% for polak-ribiere:dgg = sum((xi+g) .* xi);if (gg == 0.0) return; end;gam = dgg / gg;g = -xi;h = g + gam * h;x1 = h;end% error(`Too many iterations in frprmn`);disp(`Warning: Too many iterations in frprmn`);file: /homes/leventon/Matlab/g2x0.mfunction v = g2x0(x, y)v = 0.9213*(2*x*x-1)*exp(-(x*x+y*y));file: /homes/leventon/Matlab/gaus.mfunction v = gaus(sig)no2 = sig*4;x = -no2 : no2;v = 1/(sqrt(2*pi)*sig) * exp(-(x.*x)/(2*sig*sig));file: /homes/leventon/Matlab/gaus2.mfunction I = gaus2(width, IO)if (nargin < 2)IO = 1;endI = IO;for i = 1:widthI = conv2([1 1], [1 1], I)/2;endfile: /homes/leventon/Matlab/gausintLine.mfunction val = gausIntLine(I, p0, sigma, p1, p2, val)N = size(p1, 1);for t = 1:Np0 = p0(:);p1 = p1(:);p2 = p2(:);d1 = norm(p1-p0);d2 = norm(p2-p0);dm = norm((p1+p2)/2-p0);p1p2 = p1.*p2;p1p1 = p1. 2;p2p2 = p2. 2;p1p0mp2p0 = p1.*p0 - p2.*p0;dp1 = sign(sum(-p1p1 + p1p2 + p1p0mp2p0));dp2 = sign(sum(-p2p2 + p1p2 - p1p0mp2p0));if (dp1 == dp2)p21 = p2-p1;theta = atan2(p21(2), p21(1));c = cos(theta);s = sin(theta);pmrx = [c s] * (p0-p1);pm = [c; s] * pmrx + p1;dm = norm(pm-p0);val = abs(erf(d2/sigma)-erf(dm/sigma)) +abs(erf(dm/sigma)-erf(d1/sigma));elseval = abs (erf(d2/sigma)-erf(dm/sigma));endfile: /homes/leventon/Matlab/gausline.mfunction I = gausline(I, p1, p2, val, combineType, sigma)% gausLine Add a gaussian line to an imageif (nargin < 5)combineType = `add`;endif (nargin < 6)sigma = 10;endN = size(p1, 1);p0 = zeros(N, 2);for x = 1:size(I,1)fprintf(1, `.`);p0(:,i) = x;for y = 1:size(I,2)fprintf(1, `.`);p0(:,2) = y;d1 = sqrt(sum((p1-p0). 2, 2));d2 = sqrt(sum((p2-p0). 2, 2));p1p2 = p1.*p2;p1p1 = p1. 2;p2p2 = p2. 2;p1p0mp2p0 = p1.*p0 - p2.*p0;dp1 = sign(sum(-p1p1 + pip2 + p1p0mp2p0, 2));dp2 = sign(sum(-p2p2 + pip2 - p1p0mp2p0, 2));for t = 1:Nif (dp1(t,:) == dp2(t,:)) p21 = p2(t,:)-p1(t,:); theta = atan2(p21(2), p21(1)); c = cos(theta); s = sin(theta); pmrx = (p0(t,:)-p1(t, :)) * [c; s]; pm = [c s] * pmrx + p1(t,:); dm = norm(pm-p0(t,:)); erfdm = erf(dm/sigma); vals = val * (abs(erf(d2(t)/sigma)-erfdm) + abs(erfdm-erf(d1(t)/sigma))); I(x,y,:) = I(x,y,:) + shiftdim(vals(t,:), -1);else vals = val * abs(erf(d2(t)/sigma)-erf(d1(t)/sigma)); I(x,y,:) = I(x,y,:) + shiftdim(vals(t,:), -1);endendendendfile: /homes/leventon/Matlab/gausmix.mfunction [priors, mu, sigma] = gausmix(y, K, nIter, priors, mu, sigma)%% This code was adapted from the PS3 Solution Set[N, D] = size(y);if (nargin < 6)priors = ones(1, K);mu = ((rand(D, K)) .* (ones(K,1)*(max(y)-min(y)))' +(ones(K,1)*min(y))');sigma = repmat(eye(D), [1, 1, K]) * 20;endh = zeros(N, K);gaussian =inline('1/((2*pi) (length(sigma)/2) *prod(sigma) )*exp(-1/2*sum((x-mu*ones(1,size(x,2))). 2.*(sigma. -2* ones(1,size(x,2))),1)) ', `x`. `mu`, `sigma`);gaussian =inline('1/(2*pi) (lenmgth(sigma)/2)*abs(det(sigma)))*exp(-1/2(*sum((x-repmat(mu,1, size(x,2))) .* (inv(sigma) * (x-repmat(mu, 1, size(x,2))))))',`mu`, `sigma`);hold offplot(y(:,1), y(:,2), `.`)hold onplot(mu(1,:) , mu(2,:), `r*`)pause(0.1);for iter = 1:nIterfor k = 1:Kh(:,k) = priors(k) * gaussian(y', mu(:,k), sigma(:,:,k))';endh = h ./ (sum(h,2) * ones(1, K));priors = sum(h, 1);for k = 1:Kmu(:,k) = y' * h(:,k) / priors(k);% sigma(:,k) = sqrt((((y-ones(N,1)*mu(:,k)'). 2)' * h(:,k)) /priors(k));val = (y-ones(N,1)*mu(:,k)`)`;sigma(:,:,k) = val * (val' .* repmat(h(:,k), 1, D)) / priors(k);endpriors = priors / N;%% If all sigmas are the same, uncomment this line:% sigma = mean(sigma(:)) * ones(D, K);%% If all gaussians are uniform, (but different gaussians have diffsigmas),%% then uncomment this line:% sigma = repmat(mean(sigma), D, 1);if (iter .sup..about. = nIter)% sigma = max(sigma, 20);endif (mod(iter, 10) == 0)hold offplot(y(:,1) , y(:,2), `.`)hold onplot(mu(1,:), mu(2,:), `r*`)pause(0.1);iterpriorsmusigmaendendhold offplot(y(:,1), y(:,2), `.`)hold onplot(mu(1,:), mu(2, :), `r*`)pause(0.1);file: /homes/leventon/Matlab/gausmix.sub.-- sample.mfunction pts = gausmix.sub.-- sample(n, priors, mu, sigma)K = length(priors);D = size(mu, 1);pts = zeros(D, n);for i = 1:nr = rand;for k = 1:Kif (r < priors(k))break;elser = r - priors(k);endendpts(:, i) = sigma(:,:,k)*randn(2,1) + mu(;,k);endfile: /homes/leventon/Matlab/getSD.sub.-- no.sub.-- normals.m%% getSD.sub.-- no.sub.-- normals.m get derivatives of sensor responseswith%% respect%% to body point positions. Like getSD, except omits the%% derivatives with respect to normal directions.%% Sept. 3, 1997 freeman created and verified that it works. see%% test.m%% Sept. 4, 1997 modified to include derivatives with respect to%% point orientations.%% oct. 15, 1997 wtf removed derivs wrt point oris, for speed.%% Removing%% that calc changed compute time from 54 seconds to 3%% seconds, for particular parameter settings.%% oct. 17, 1997 made pt.sub.-- ori.sub.-- mod, exp.sub.-- modulated global variables, not%% passed.%% inputs: sensor.sub.-- pos array of sensor positions. 1st row: all x%% positions;%% 2nd row: all y%% positions;%% num.sub.-- ori number of orientation tunings for the%% sensors.%% pt.sub.-- pos array of (body) point positions. Same row%% layout as sensor.sub.-- pos.%% pt.sub.-- ori.sub.-- mod array of modulations based on body point%% orientations.%% These are respresented as multipliers for the%% integration strength for each of the num.sub.-- ori%%orientations.%% Thus, pt.sub.-- ori has num.sub.-- ori rows, and as many%% columns as pt.sub.-- pos does. Calculated in%%getsR.m%% twosigmasq 2 * sigma 2 for the receptive field%% fall-offs.%% outputs: derivatives of the responses with respect to the pt.sub.-- pos%% and pt.sub.-- ori values.%% derivs is a matrix: num.sub.-- sensors * num.sub.-- ori columns by%% num.sub.-- pts * 4 (for x and y%% components of both position and%% orientation)%% The ordering of the rows:%% 1st row: deriv w.r.t. x component of pt 1.%% 2nd row: deriv w.r.t. y component of pt 1.%% 3rd row: deriv w.r.t. x component of pt 2.%% . . .%% See also the function getSR.m%% Note: the computation of the orientation derivatives can be%% speeded-up by using repmat like operations, instead of doing the%% looping that's done here. Sept. 5, 1997 wtf.function [derivs] = getSD no.sub.-- normals(sensor.sub.-- pos,sensor.sub.-- ori, pt.sub.-- pos, pt.sub.-- ori, . . . twosigmasq, cos.sub.-- power);global pt.sub.-- ori.sub.-- mod exp.sub.-- modulated;[num.sub.-- ori, num.sub.-- pts] = size(pt.sub.-- ori.sub.-- mod);[tmp, num.sub.-- sensors] = size(sensor.sub.-- pos);%% pt.sub.-- sensor contains (x.sub.-- point - x.sub.-- sensor) and(y.sub.-- point - y.sub.-- sensor)pt.sub.-- sensor =repmat( . . . reshape(repmat(reshape(sensor.sub.-- pos', 1, 2*num.sub.-- sensors), num.sub.-- ori, 1) . . . , num.sub.-- ori * num.sub.-- sensors, 2)', . .. num.sub.-- pts, 1) - . . . repmat( reshape(pt.sub.-- pos, num.sub.-- pts*2, 1), 1, num.sub.-- sensors*num.sub.-- ori);%% nowform an array of (x.sub.-- point - x.sub.-- sensor) 2 + (y.sub.--point - y.sub.-- sensor) 2tmp.sub.-- arr = pt.sub.-- sensor) 2;tmp.sub.-- arr = tmp.sub.-- arr + tmp.sub.-- arr([2 : (2*num.sub.--pts),1], :);tmp.sub.-- arr(2:2:(2*num.sub.-- pts), :) = tmp.sub.-- arr(1:2:(2*num.sub.-- pts), :);%% the derivatives with respect to point x and y positionsderivs = exp(-(tmp.sub.-- arr)/twosigmasq) .* . . .pt.sub.-- sensor .* . . . repmat(reshape(repmat(reshape(pt.sub.-- ori.sub.-- mod',1,num.sub.-- pts*num.sub.-- ori), . . . 2,1), 2*num.sub.-- pts, num.sub.-- ori), 1 num.sub.-- sensors) . . . * 2 / twosigmasq;file: /homes/leventon/Matlab/getSR.m%% getSR.m get sensor responses.%% Simulates what the response would be to a particular%% set of point positions.%% For now: input is points on oriented segments, and their%% orientations.%% Later: modify to let the inputs be coefficient values.%% Sept. 4, 1997 freeman created.%% oct. 17, 1997 made pt.sub.-- ori.sub.-- mod, exp.sub.-- modulated global variables, not%%%% passed.%% inputs: sensor.sub.-- pos array of sensor positions. 1st row: all x%% positions;%% 2nd row: all y%% positions;%% sensor.sub.-- ori array of the normal vectors for each%% orientation that all the sensors take on.%% pt.sub.-- pos Array of (body) point positions. Same row%% layout as sensor.sub.-- pos.%% pt.sub.-- ori Array of normal vectors for each point. Each%%%% column is a normal vector of a point.%% twosigmasq 2 * sigma 2 for the receptive field%% fail-offs.%% cos.sub.-- power Fall-off factor for the orientation tuning.%% It goes as cos(theta1-theta2) cos.sub.-- power.%%%% outputs: response Array of sensor responses, each row at one of%% the orientations%% (2-d array), each column at a different%% sensor position.%% pt.sub.-- ori.sub.-- mod Array of the sensor modulation as a function%% of orientation of each of the points. Used%% in calculating%% derivatives, so output it so we don't have to%% recompute it.%% exp.sub.-- modulated The response modulations between each%% sensor and each point, due to distance%% variations. Again, for the%% derivative calculation we don't want to%% have to recompute this.function [response] = getSR(sensor.sub.-- pos, sensor.sub.-- ori,pt.sub.-- pos, pt.sub.-- ori, . . . twosigmasq, cos.sub.-- power); global pt.sub.-- ori.sub.-- mod exp.sub.-- modulated; [tmp, num.sub.-- sensors = size(sensor.sub.-- pos); [tmp, num.sub.-- ori] = size(sensor.sub.-- ori); [tmp, num.sub.-- pts] = size(pt.sub.-- pos); response = zeros(num.sub.-- ori, num.sub.-- sensors);dist.sub.-- sqr = ( repmat( sensor.sub.-- pos(1,:), num.sub.-- pts,1) - . . . repmat( pt.sub.-- pos(1,:)', 1,num.sub.-- sensors)) . 2 + . . . ( repmat( sensor.sub.-- pos(2,:), num.sub.-- pts,1) - . . . repmat( pt.sub.-- pos(2,:)', 1,num.sub.-- sensors)) . 2;%% exp.sub.-- modulated is the distance modulated response of ever sensorto%% every point of the figure. Doesn't yet take orientation into account.exp.sub.-- modulated = exp(- dist.sub.-- sqr / twosigmasq);%% an array of the modulation based on orientation tuning of each%% sensor.pt.sub.-- ori.sub.-- mnod = sensor.sub.-- ori' * pt.sub.-- ori;pt.sub.-- ori.sub.-- mod = pt.sub.-- ori.sub.-- mod. cos.sub.-- power;%% fix annoying non-consistant behavior of "sum" for num.sub.-- pts = 1and%% otherwise.for ori = 1:num.sub.-- oriif (num.sub.-- pts == 1)response(ori, :) = repmat(pt.sub.-- ori.sub.-- mod(ori,:)', 1, num.sub.--sensors) .*exp.sub.-- modulated;elseresponse(ori, :) = . . . sum(repmat(pt.sub.-- ori.sub.-- mod(ori,:)', 1, num.sub.-- sensors) .*exp.sub.-- modulated);endendfile: /homes/leventon/Matlab/grad.mfunction [xR, yT, R, T] = grad(I, ang)[m, n] = size(I);x = I(2:m,1:n-1)-I(1:m-1,1:n-1);y = I(1:m-1,2:n)-I(1:m-1,1:n-1);R = x.*x + y.*y;T = atan2(y,x);if (nargin == 1)xR = x;yT = y;elsexR = R;yT = T;endfile: /homes/leventon/Matlab/h2x0.mfunction v = h2x0(x, y)v = (-2.205*x + 0.9780*(x 3))*exp(-(x*x+y*y));file: /homes/leventon/Matlab/halfsize.mfunction Mh = halfsize(M)[m, n] = size(M);Mh = (M(1:2:m-1, 1:2:n-1) + M((1:2:m-1)+1, (1:2:n-1)+1))/2;file: /homes/leventon/Matlab/haus.mfunction [match, fbest] = haus(model, im, varargin)x0 = 1;y0 = 1;xm = size(im, 1)-size(model,1);ym = size(im, 2)-size(model,2);argi = 1;argc = nargin - 2;while (argi <= argc)switch varargin{argi}case `r`x0 = max(x0, varargin{argi+i});y0 = llLax(y0, varargin{argi+2});xm = min(xm, varargin{argi+3});ym = min(ym, varargin{argi+4});argi = argi + 5;otherwiseerror(`Usage Error`);endendmodel = uint8(model.sup..about. =1);im = dilate(uint8(im.sup..about. =i), ones(3));% im = dilate(uint8(im.sup..about. =1), [0 1 0; 1 1 1; 0 1 0]);% im = dilate(uint8(im.sup..about. =1), [0 0 1 0 0; 0 1 1 1 0; 1 1 1 1 1;0 1 1 1% 0; 0 0 1 0 0]);% [modelptsX, modelptsY] = ind2sub(size(model), find(model));% modelpts = [modelptsX modelptsY];dx = 1;dy = 1;vbest = -1;for y = y0:dy:ymys = y:y+size(model, 2)-1;for x = x0:dx:xmxs = x:x+size(model, 1)-1;v = length(find(and(model, im(xs, ys))));if (v >= vbest)vbest = v;match = [x y];endxs = xs + 1;endys = ys + 1;endfbest = vbest / length(find(model));file: /homes/leventon/Matlab/headpos.mfunction hpos = headpos(seq)% headpos Compute the head position of a sequence.global seqINFOseq = seqconvert(seq, `ang`);data = seq.data;model = seqINFO.repAngle.model;T = size(data, 2);hpos = zeros(T, 3);for t = 1:Tmodel = setModelData(model, data(:,t));parts = model{8};pos = zeros(1,6);for i = 2:7if (isempty(model{i}) == 0)pos(i-1) = model{i};endendn = size(parts, 2);part = parts{1};pg = pos(1:3);og = pos(4:6);pos1 = zeros(1:6);for i = 5:10if (isempty(part{i}) == 0)pos1(i-4) = part{i};endendpl = pos1(1:3);ol = pos1(4:6);Rg = aToR(og);R1 = aToR(ol);Tg = [[Rg; 0 0 0] [diag(diag(pg)); 1]];T1 = [[R1; 0 0 0] [diag(diag(pl)); i]]T = Tg * T1;o = rToA(T(1:3,1:3));p = T(1:3,4)d = T(1:3,4)' T(1:3,3)'];rx = part{2};ry = part{3};rz = part{4};T2 = T * [1 0 0 0; 0 ry/rx 0 0; 0 0 1 0; 0 0 0 1];d = [T2(1:3,1) T2(1:3,2) T2(1:3,3) T2(1:3,4)];subparts = part{11};part = subpafts{3};pg = p;og = o;pos1 = zeros(1:6);for i = 5:10if (isempty(part{i}) == 0)pos1(i-4) = part{i};endendpl = pos1(1:3);ol = pos1(4:6);Rg = aToR(og);Rl1 = aToR(ol);Tg = [[Rg; 0 0 0] [diag(diag(pg)); 1]];T1 = [[R1; 0 0 0] [diag(diag(pi)); 1]];T = Tg * T1;o = rToA(T(1:3,1:3));p = T(1:3,4)'d = [T(1:3,4)' T(1:3,3)'];rx = part{2};ry = part{3};rz = part{4};T2 = T * [1 0 0 0; 0 ry/rx 0 0; 0 0 1 0; 0 0 0 1];d = [T2(1:3,1) T2(1:3,2) T2(1:3,3) T2(1:3,4)];hpos(t, :) = d(10:12);endfile: /homes/leventon/Matlab/imCrop.mfunction Ic = imcrop(I, x, y, w, h)[m, n, p] = size(I);if (p > 1)Ic = zeros(w, h, p);for i = i:pIc(:,:,i) = imCrop(I(:,:,i), x, y, w, h);endreturnendleft = [ ];right = [ ];top = [ ];bottom = [ ];yof = 0;if (y < 1)left = zeros(w, -y);yof = -y;h = h +y;y = 1;endif (h > n-y+1)h2 = n-y+1;right = zeros(w, h-h2);h = h2;endif (x < 1)top = zeros(-x, h);w = w + x;x = 1;endif (w > m-x+1)w2 = m-x+1;bottom = zeros(w-w2, h);w = w2;endIc = [left [top; J(x:x+w-1, y:y+h-1); bottom] right];file: /homes/leventon/Matlab/imDiff.mfunction C = imDiff(A, B, thresh)[m, n] = size(A);A = A(2:m-1, 2:n-1);Cul = abs(A-B(1:m-2, 1:n-2));Cuc = abs(A-B(2:m-1, 1:n-2));Cur = abs(A-B(3:m, 1:n-2));Ccl = abs(A-B(1:m-2, 2:n-1));Ccc = abs(A-B(2:m-1, 2:n-1));Ccr = abs(A-B(3:m, 2:n-1));Cll = abs(A-B(1:m-2, 3:n));Clc = abs(A-B(2:m-1, 3:n));Clr = abs(A-B(3:m, 3:n));CC(:,:,1) = Cul;CC(:,:,2) = Cuc;CC(:,:,3) = Cur;CC(:,:,4) = Ccl;CC(:,:,5) = Ccc;CC(:,:,6) = Ccr;CC(:,:,7) = Cll;CC(:,:,8) = Clc;CC(:,:,9) = Clr;C = min(CC, [ ], 3);% C = ((Cul > thresh) & (Cuc > thresh) & (Cur > thresh) &% (Ccl > thresh) & (Ccc > thresh) & (Ccr >thresh) &% (Cll > thresh) & (Clc > thresh) & (Clr > thresh));C = [zeros(m, 1) [zeros(1, n-2); C; zeros(1,n-2)] zeros(m, 1)];mask = [0 0 1 0 0; 0 1 1 1 0; 1 1 1 1 1; 0 1 1 1 0; 0 0 1 0 0]% C = double(dilate(C, mask));file: /homes/leventon/Matlab/imDiff2.mfunction C = imDiff(A, B)[m, n] = size(A);D = abs(A-B);C = D.*Afile: /homes/leventon/Matlab/imRotate.mfunction Ir = imRotate(I, theta)[m, n] = size(I);Ir = zeros(m,n);c = cos(theta);s = sin(theta);R = [c s; -s c];mh = round(m/2);nh = round(n/2);for i=i:mfor j=1:npr = R*[(i - mh); (j - nh)] + [mh; nh];prf = floor(pr);dpr = pr-prf;dpr = pr-prf;if ((min(prf) < 1) .vertline. (prf(1) > m-1) .vertline. (prf(2) > n-1))Ir(i,j) = 0;elsev = I(prf(1), prf(2))*(1-dpr(1))*(1-dpr(2));v = v + I(prf(1)+1, prf(2))*(dpr(1))*(1-dpr(2));v = v + I(prf(1), prf(2)+1)*(1-dpr(1))*(dpr(2));v = v + I(prf(1)+1, prf(2)+1)*(dpr(1))*(dpr(2));Ir(i,j) = v;endend;end;file: /homes/leventon/Matlab/imSelect.mfunction [I, p = imSelect(I, p)if (nargin == 2)I = I(p(1,1):p(2,1), p(1,2):p(2,2));elsehold offimagesc(I);axis imagep = round(ginput(2));if (isempty(find(p<=0)) == 0)I = [ ];p = [ ];returnendp = p(:, [2 1]);I = I(p(1,1):p(2,1), p(1,2):p(2,2));endfile: /homes/leventon/Matlab/initSensors.m%% Where should these initializations go? wtf oct. 21 97%% set-up array of sensorssensor.sub.-- rows = 8;sensor.sub.-- cols = 10;minx = -30;maxx = 30;miny = -10;maxy = 30;delx = (maxx - minx) / (sensor.sub.-- rows - 1);dely = (maxy - miny) / (sensor.sub.-- cols 3- 1);%% evenly spaced sensors[tmp.sub.-- x, tmp.sub.-- y] = meshgrid(minx:delx:maxx, miny:dely:maxy);sensor.sub.-- pos = [reshape(tmp.sub.-- x, 1, sensor.sub.-- rows*sensor.sub.-- cols); . . .reshape(tmp.sub.-- y, 1, sensor.sub.-- rows*sensor.sub.-- cols)];twosigmasq = 2 * 4 2;cos.sub.-- power = 0;num.sub.-- ori = 1;%% evenly spaced between 0 and 180. For one ori, just fillers.tmp.sub.-- ori = (0:(num.sub.-- ori-1))*pi/num.sub.-- ori;sensor.sub.-- ori = [cos(tmp.sub.-- ori); sin(tmp.sub.-- ori)];file: /homes/leventon/Matlab/isKnowledge.mfunction res = isKnowledge(a)if (isstruct(a) == 0)res = 0;returnendf = fields(a);res = 0;for i = 1:size(f,1)if (strcmp(f(i), `enc`))res = 1;returnendendfile: /homes/leventon/Matlab/isseqType.mfunction res = isSeqType(a)res = 0;if (isstr(a) == 0)return;endswitch acase {`3d`, `2d`, `head`, `motion`, `image`, `form`, `ang`}res = 1;endfile: /homes/leventon/Matlab/isSequence.mfunction res = isSequence(a)if (isstruct(a) == 0)res = 0;returnendf = fields(a);res = 0;for i = 1:size(f,1)if (strcmp(f(i), `nframes`))res = 1;returnendendfile: /homes/leventon/Matlab/kth.mfunction v = kth(mx, num)b = (num-1)*mx+1;v = b:b+mx-1;file: /homes/leventon/Matlab/laplacian.mfunction I = laplacian(I)mask = [1 4 1; 4 -20 4; 1 4 1];I = conv2(I, mask, `same`);file: /homes/leventon/Matlab/linmin.m%% linmin.m p. 419 num rec in c.%% do multi-dimensional function minimization along a given high-d%% line.%% input variables:%% p: high dixnensional point. xi: direction along which to search.%% func: multi-dimensionai function to look at.%% output variables:%% p: where func(p) takes a miniuium along the line.%% xi: the actual vector displacement between start and min points p.function [p, xi, fret] = linmin(p, xi, func)global pcom xicom nrfunc;TOL = 2.0e-4;nrfunc = func;pcom = p;xicom = xi;ax = 0; xx = 1.0;[ax, xx, bx, fa, fx, fb] = mnbrak(ax, xx, `fldim`);[fret, xmin] = brent(ax, xx, bx, `fldim`, TOL);xi = xmin * xi;p = p + xi;file: /homes/leventon/Matlab/maddscale.mfunction C = maddscale(A, sa, B, sb)[n, m] = size(A);C = [A(1,:); (sa * A(2:n,:) + sb * B(2:n,:))];file: /homes/leventon/Matlab/make.sub.-- stick.sub.-- figs.m%% make.sub.-- stick.sub.-- figs.m%% Generates figures for cvpr paper. For the stick figure part.%% Nov. 22, 1997%% nov. 25, 1997 commented out all the postscript file print%% statements, to not overwrite things I want . . .fig = 3;%, InitializeseqInitif (fig == 1)% display several of the sequences.seq.sub.-- list = [2,3,4];frames.sub.-- list = {[1:20:200], [1:20:200], [1:15:200], [1:6:100],[1:10:100]};offset.sub.-- list = {[0 -20 0], [0 -30 0], [-20 -10 0], [-10 0 0], [-100]};view.sub.-- list = {[-30,30], [-30,10], [-30,30], [-30,30], [-30,30]};i.sub.-- plot = 0;for i = seq.sub.-- listi.sub.-- plot = i.sub.-- plot + 1;figure(i.sub.-- plot);seq = seqGet(i);fprintf(1, `seq: %d n frames: %d.backslash.n`, i, seq.nframes);seqView(seq,frames.sub.-- list{i},`stopani`,1,`expand`, offset.sub.--list{i},. . . `viewang`, view.sub.-- list{i});call = [`print /homes/freeman/ps/body/cvpr.train` num2str(i) `.ps-dpsc`];%%% eval(call);endend % fig == 1if (fig == 2)%%% train on all sequences, rotated.train.sub.-- set = [1:10];for num.sub.-- angles = 2angle.sub.-- set = [0:pi/num.sub.-- angles:pi-pi/num.sub.-- angles];sel = seqSelect(train.sub.-- set, angle.sub.-- set};knl3d = seqTrain(sel, `3d`, `submean`, 1, `MtM`, 0);%% The encoded training matriciesM3 = knl3d.enc{i};mean3 = knl3d.mean{i};%% Taking the SVD of these training matriciesfprintf(1,`pre svd, num.sub.-- angles = %d.backslash.n`, num.sub.--angles);[U3, S3, V3] = svd(M3, 0);fprintf(1,`post svd, num.sub.-- angles = %d.backslash.n`, num.sub.--angles);semilogy(diag(53));hold onendhold offtmp = diag(S3);save /homes/freeman/matlab/body/s3diag tmpnEv = 50;obs2d = seqGet(3, `2d`, pi/2);frames = [1:2:10];offset = [10 0 0];%% show the gaussian drawsfor trial = 1:3alpha = sqrt(diag(S3(1:nEv,1:nEv))) .* randn(nEv,1);seqView(seqDecode(mean3 + U3(:,1:nEv) * alpha, `3d`, . . .knl3d.encParams, obs2d), frames,`stopani`,1, . . . `expand`, offset);call = [`print /homes/freeman/ps/body/randn.pca` num2str(trial) `.ps-dpsc`];%%% eval(call);end%% try the random draw, from no multiple rotation training, just 1.angle.sub.-- set = pi/2;sel = seqSelect(train.sub.-- set, angle.sub.-- set);knl3d1 = seqTrain(sel, `3d`, `submean`, 1, `MtM`, 0);M31 = knl3d1.enc{1};mean31 = knl3d1.mean{1};%% Taking the SVD of these training matricies[U31, S31, V31] = svd(M31, 0);%% show the gaussian drawsfor trial = 1:3alpha1 = sqrt(diag(S31(1:nEv,1:nEv))) .* randn(nEv,1);seqView(seqDecode(mean31 + U31(:,1:nEv) * alpha1, `3d`, . . .knl3d.encParams, obs2d), frames,`stopani`,1, . . . `expand`, offset);call = [`print /homes/freeman/ps/body/randni.pca` num2str(trial)`.ps -dpsc`];%%% eval(call);endend % fig == 2%% ***********************************************%% Code below is to make figures 5 and 6 of the cvpr submission.% Train on all but one sequencesel.sub.-- index = 3;%% all but sel.sub.-- index. Works for 1 and 10, too.train.sub.-- set = [1:(sel.sub.-- index-1) (sel.sub.-- index+1):10];sel = seqSelect(train.sub.-- set, pi/2);knl3d = seqTrain(sel, `3d`);knl2d = seqTrain(sel, `2d`);%% The encoded training matriciesM3 = knl3d.enc{1};%% Get the observationobs2d = seqGet(sel.sub.-- index, `2d`, pi/2);obs3d = seqGet(sel.sub.-- index, `3d`, pi/2);obs2dc = seqcrop(obs2d, 1, 22);obs3dc = seqcrop(obs3d, 1, 22);%% Encode the observation (3 snippets worth)enc2dc = seqEncode(obs2dc);%% Let y be just the first snippet of the observationy = enc2dc(:,1);%% Make a projection matrix that takes a matrix%% from (3*370 .times. N) to (2*370 .times. N)P = project.sub.-- matrix(370);%% Set numbers of E-vectsnEv = 50;%% the true sequencefigure(1)expand.sub.-- vec = [3 0 0];seqView(obs3dc, [ ]);seqView(obs3dc, [ ], `filespec`, . . . `/homes/freeman/images/body/tmp.sub.-- for.sub.-- cvpr/seq5a-%03d. ps`);/ -%%% seqView(obs3dc, [ ], `stopani`, 1, `expand`, expand.sub.-- vec);%%% print /homes/freeman/ps/bodylorigseq.ps -dpsc;%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% Try the kalman filter MAP estimate of the 3d.%% (This should be very similar (identical?) to the "short path",%% with a prior term included of the proper weight).%% Work with mean zero training matrix. Just remember to add back%% in the mean at the end.mean3 = mean(M3, 2);M3zeromean = M3 - repmat(mean3, 1, size(M3,2));[U30, S30, V30] = svd(M3zeromean, 0);%% do the kalman filter reconstruction (bayesian).%% From wtf composition book,obs.sub.-- noise = 100;g = S30(1:nEv,1:nEv) * U30(:,i:nEv)' * P` * . . . inv(p*U30(:,1:nEv)*S30(1:nEv,1:nEv)*U30(:,1:nEv)'* P`. . . + obs.sub.-- noise*eye(size(y,1)));%% modify the observations, enc2dc, to be zero mean, too.alphak = g * (enc2dc - repmat(P * mean3, 1, size(enc2dc,2)));figure(2)seq = seqDecode(repmat(mean3,1,size(enc2dc,2)) + . . . U30(:,1:nEv) * alphak(1:nEv,:), `3d`, . . .knl2d.encParams, obs2d);seqview(seq, [ ]);seqview(seq, [ ], `filespec`, . . . `/homes/freeman/images/body/tmp.sub.-- for.sub.-- cvpr/seq5d-%03d. ps`);%% seqView(seqDecode(repmat(mean3,1,size(enc2dc, 2)) + . . .%% U30(:,1:nEv) * alphak(1:nEv,:), `3d`, . . .%% knl2d.encParams, obs2d), [ ], `stopani`, 1, `expand`,%% expand.sub.-- vec);%%% print /homes/freeman/ps/body/kal.100.ps -dpsc;fprintf(1, `kalman fit to 2d data: %f.backslash.n`, . . . norm(enc2dc(:,1) - . . . P * (mean3 + U30(:,1:nEv) * alphak(1:nEv, 1))));%% The covariance matrix of the posterior distribution of the%% 3-d figure. From the standard bayesian/kalman filter completing%% the square. This posterior covariance should tell you how much%% uncertainty there is about the figure after the measurement, and%% in what directions it is.postCov = S30(1:nEv,1:nEv) - . . . g * P * U30(:,1:nEv) * S30(1:nEv,1:nEv);figure (4)semilogy(diag(S30(1:nEv, 1:nEv)));v = axis; v(2) = 50; v(3) = 10; v(4) = 100000;hold on;semilogy(diag(postcov), `--`);axis(v);hold off;title(`prior: solid posterior: dashed`);%%% print /homes/freeman/ps/body/svd.ps -dpsc;%% Dominated by contribution in mode U30(:,2)%% Display that mode: (or else the next 2 highest: 1 and 10).figure (5)%% we don't know how he's sliding;seq = seqDecode(-10 * U30(:,2) , `3d`, . . .knl2d.encParams, obs2d);seqView(seq, [ ]);seqView(seq, [ ], `filespec`, . . . `/homes/freeman/images/body/tmp.sub.-- for.sub.-- cvpr/seq6c-%03d. ps`);%% seqView(seqDecode(U30(:,10) , `3d`, . . .%% knl2d.encParams, obs2d), [ ], `stopani`, 1, `expand`, [0 00];%%% print /homes/freeman/ps/body/mode2.ps -dpsc;figure (6)we don't know where he's standingseq = seqDecode(10 * U30(:,i) , `3d`, . . .%% knl2d.encParams, obs2d),seqView(seq, [ ]);seqView(seq, [ ], `filespec`, . . . `/homes/freeman/images/body/tmp.sub.-- for.sub.-- cvpr/seq6b-%03d. ps`);%% seqView(seqDecode(10 * U30(:,1) , `3d`, . . .%% knl2d.encParams, obs2d), [ ], `stopani`, 1, `expand`, [0 00]);%%% print /homes/freeman/ps/body/model.ps -dpsc;figure(7)we don't know how he's moving his arms towards the cameraseq = seqDecode(U30(:,10) , `3d`, . . .knl2d.encParams, obs2d);seqView(seq, [ ]);seqView(seq, [ ], `filespec`, . . . `/homes/freeman/images/body/tmp.sub.-- for.sub.-- cvpr/seq6c-%03d. ps`);%% seqView(seqDecode(U30(:,10) , `3d`, . . .%% kn12d.encParams, obs2d), [ ], `stopani`, 1, `expand`, [0 00]);%%% print /homes/freeman/ps/body/mode10.ps -dpsc;%% the 2-d image of the true sequencefigure (8)expand.sub.-- vec = [3 0 0];seqView(obs3dc, [ ]);seqView(obs3dc, [ ], `filespec`, . . . `/homes/freeman/images/body/tmp.sub.-- for.sub.-- cvpr/seq5b-%03d. ps`, . . . `viewang`, [0 0], `axisval`, [-20, 80, 0, 20,0, 70] );%% seqView(obs3dc, [ ], `stopani`, 1, `expand`, expand.sub.-- vec, . . .%% `viewang`, [0 0], `axisval`, [-20, 80, 0, 20,0, 70]);%% print /homes/freeman/ps/body/origseq2d.ps -dpsc;%% do the kalman filter reconstruction with virtually no prior infoobs.sub.-- noise= 0.0001;g = S30(1:nEv,1:nEv) * U30(:,1:nEv)'* P' * . . . inv(P*U30(:,1:nEv)*S30(1:nEv,1:nEv)*U30(:,1:nEv)'* P'. . . + obs.sub.-- noise*eye(size(y,1)));%% modify the observations, enc2dc, to be zero mean, too.alphak = g * (enc2dc - repmat(P * mean3, 1, size(enc2dc,2)));figure (3)seq = seqDecode(repmat(mean3,1,size(enc2dc,2)) + . . . U30(:,1:nEv) * alphak(1:nEv,:), `3d`, . . .knl2d.encParams, obs2d);seqView(seq, [ ];seqView(seq, [ ], `filespec`, . . . `/homes/freeman/images/body/tmp.sub.-- for.sub.`3 cvpr/seq5c-%03d. ps`)%% seqView(seqDecode(repmat(mean3,1,size(enc2dc,2)) + . . .%% U30(:,1:nEv) * alphak(1:nEv,:), `3d`, . . .%% knl2d.encParams, obs2d), [ ], `stopani`, 1, `expand`, expand.sub.-- vec);%%% print /homes/freeman/ps/body/kal.0001.ps -dpsc;%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% print out the covariance numbers%% volume of the gaussians: proportional to sqrt of det of cov mat.pri = sqrt(det(S30(1:nEv,1:nEv)));pos = sqrt(det(postCov));fprintf(1,`prior volume: %e posterior volume: %e.backslash.ratio:%e.backslash.n`, . . . pri, pos, pos/pri);%% prior volume: 3.173442e+65 posterior volume: 1.971117e+52%% ratio: 6.211292e-14pri2 = sqrt(det(S30(3:nEv,3:nEv)));pos2 = sqrt(det(postcov(3:nEv, 3:nEv)));fprintf(1,`excluding those rigid modes: .backslash.nprior volume: %eposterior%volume:%%e.backslash.nratio:%%%e.backslash.n', . . . pri2, pos2, pos2/pri2);%% excluding those rigid modes:%% prior volume: 3.318102e+61 posterior volume: 3.602475e+49%% ratio: 1.085704e-12%% ask: what percent of the variance is explained by the 1st 50%% eigenvalues??sum50 = sum(diag(S30(1:nEv, 1:nEv)));sumall = sum(diag(S30));fprintf(1,`sum50: %f, sumall: %f, ratio: %f.backslash.n`, . . . sum50, sumall, sum50/sumall);%% sum50: 45535.571955, sumall: 49956.542129, ratio: 0.911504%% hmm how about how much energy is in the two signals, prior and%% posterior?sumpos = sum(diag(postcov(1:nEv, 1:nEv)));fprintf(1,`sum50: %f, sumpos: %f, ratio: %f.backslash.n`, . . . sum50, sumpos, sum50/sumpos);sum502 = sum(diag(S30(3:nEv, 3:nEv)));sumpos2 = sum(diag(postcov(3:nEv, 3:nEv)));fprintf(1,`sum502: %f, sumpos2: %f, ratio: %f.backslash.n`, . . . sum502, sumpos2, sum502/sumpos2);%% sum50: 45535.571955, sumpos: 15312.857358, ratio: 2.973682%% sum502: 25796.450325, sumpos2: 8314.591388, ratio: 3.102552return;file: /homes/leventon/Matlab/makecyl.mfunction cyl = makecyl(desc, pts, r)%%%%cyl = cell(1,10);cyl{1} = desc;[n, lose] = size(pts);if (n == 2)cyl{2} = r;cyl{3} = r;cyl{4} = norm(p1-p2);cyl{5:7} = num2cell(pts(1,1:3));cyl{file: /homes/leventon/Matiab/makefigsl.mstartload plotdatafigure(1)hoid offsemilogx(logspace(-1, 1, 21), wave2(TimeScErr(:, :, 1), AoaSizes),`x-`);xlabel(`Time scale factor`);ylabel(`Average 2D Error`)title(`Average 2D Error vs. Time Scale`)figure(2),hold offsemilogx(logspace(-1, 1, 21), wave2(TimeScErr(:, :, 2), AoaSizes),`x-`)xlabel(`Time scale factor`)ylabel(`Average 3D Error`)title(`Average 3D Error vs. Time Scale')figure(3);hold offplot(s2all0d);hold on;plot(s2all0s2d, `r`);plot(s2all0s4d, `g`);plot(s2all0s8d, `k`);plot(s2all0n90d, `m`);axis([0 100 0 1000]);legend(`Train every 360 degrees`, `Train every 180 degrees`, `Trainevery 90 degrees`, `Train every 45 degrees`, `Train on 0 and 90degrees`);xlabel(`Eigen vector axis index`);ylabel(`Length of axis`)title(`Eigen vector axes for training at different views`)figure(4);hold off;plot(linspace(-180, 179, 72), wave(ViewErr(:, :, 1, 1), AoaSizes, 2));axis([-180 180 0.7 1.5]);hold onplot(linspace(-180, 179, 72), wave(ViewErr(:, :, 1, 2), AoaSizes, 2),`r`);plot(linspace(-180, 179, 72), wave(ViewErr(:, :, 1, 3), AoaSizes, 2),`g`);plot(linspace(-180, 179, 72), wave(ViewErr(:, :, 1, 4), AoaSizes, 2),`k-.`);xlabel(`Rotation (degrees)`);ylabel(`Average 2D Error (in)`);legend(`Train every 360 degrees`, `Train every 180 degrees`, `Trainevery 90 degrees`, `Train every 45 degrees`);title(`Average 2D Error vs. Rotation from trained data`)figure(5);hold off;plot(linspace(-180, 179, 72), wave(ViewErr(:, :, 1, 1), AoaSizes, 2));axis([-180 180 0.7 1.5]);hold onplot(linspace(-180, 179, 72), wave(ViewErr(:, :, 1, 2), AoaSizes, 2),`r`);plot(linspace(-180, 179, 72), wave(ViewErr(:, :, 1, 3), AoaSizes, 2),`g`);plot(linspace(-180, 179, 72), wave(ViewErr(:, :, 1, 4), AoaSizes, 2),`k-.`);plot(linspace(-180, 179, 72), wave(ViewErr(:, :, 1, 5), AoaSizes, 2),`m`);xlabel(`Rotation (degrees)`);ylabel(`Average 2D Error (in)`);legend(`Train every 360 degrees`, `Train every 180 degrees`, `Trainevery 90 degrees`, `Train every 45 degrees`, `Train on 0 and 90degrees`);title(`Average 2D Error vs. Rotation from trained data`)figure(6);hold off;plot(linspace(-180, 179, 72), wave(ViewErr(:, :, 2, 1), AoaSizes, 2));axis([-180 180 4 8]);hold onplot(linspace(-180, 179, 72), wave(ViewErr(:, :, 2, 2), AoaSizes, 2),`r`);plot(linspace(-180, 179, 72), wave(ViewErr(:, :, 2, 3), AoaSizes, 2),`g`);plot(linspace(-180, 179, 72), wave(ViewErr(:, :, 2, 4), AoaSizes, 2),`k-.`);xlabel(`Rotation (degrees)`);ylabel(`Average 3D Error (in)`);legend(`Train every 360 degrees`, `Train every 180 degrees`, `Trainevery 90 degrees`, `Train every 45 degrees`);title(`Average 3D Error vs. Rotation from trained data`)figure(7);hold off;plot(linspace(-180, 179, 72), wave(ViewErr(:, :, 2, 1), AoaSizes, 2));axis([-180 180 4 8]);hold onplot(linspace(-180, 179, 72), wave(ViewErr(:, :, 2, 2), AoaSizes, 2),`r`);plot(linspace(-180, 179, 72), wave(ViewErr(:, :, 2, 3), AoaSizes, 2),`g`);plot(linspace(-180, 179, 72), wave(ViewErr(:, :, 2, 4), AoaSizes, 2),`k-.`);plot(linspace(-180, 179, 72), wave(ViewErr(:, :, 2, 5), AoaSizes, 2),`m`);xlabel(`Rotation (degrees)`);ylabel(`Average 3D Error (in)`);legend(`Train every 360 degrees`, `Train every 180 degrees`, `Trainevery 90 degrees`, `Train every 45 degrees`, `Train on 0 and 90degrees`);title(`Average 3D Error vs. Rotation from trained data`)figure(8);hold off;plot(linspace(0,20,41), wave(NoiseScErr(:, :, 1), AoaSizes, 2), `x-`)xlabel(`Sigma (in)`);ylabel(`2D Error (in)`);title(`2D Error vs Amount of Noise`);figure (9)hold off;plot(linspace(0,20,41), wave(NoiseScErr(:, :, 2), AoaSizes, 2), `x-`)xlabel(`Sigma (in)`);ylabel(`3D Error (in)`);title(`3D Error vs Amount of Noise`);file: /homes/leventon/Matlab/makepacman.mx0 = linspace(0, 100);y0 = linspace(0, 0);k = pi/1.2;a = pi/4 * ((sin(k*(x0/100)*(2*pi)) + 1) / 2);a = pi * sin(k*(x0/100)*(2*pi));a = piI4 * (sin(k*(x0/100)*(2*pi))) + pi;for i=1:100while (a(i) > pi)a(i) = a(i) - 2*pi;endend% a = pi/4 * ((sin(k*(x0/100)*(2*pi)) + 1) / 2);x1 = x0 + 5 * cos(a);x2 = x0 + 5 * cos(a);y1 = y0 + 5 * sin(a);y2 = y0 - 5 * sin(a);Ac = [x0; y0; x1; y1; x2; y2];Aa = [x0; y0; a];for i = 1:00hold offplot([x1(i) x0(i) x2(i)], [y1(i) y0(i) y2(i)]);hold onplot([x1(i) x0(i) x2(i)], [y1(i) y0(i) y2(i)], `o`);axis([0 100 -50 50]);pause(0.1);endk = 5;s = 10;ndeg = 6;nRows = ndeg * s;nfr = 100;c = 1;for i = i:k:(nfr-s+1)v = vectorize(Ac(:, i:i+s-i)');Mc(1:nRows,c) = v;c = c + 1;end[uc, sc, vc] = svd(Mc);if (1)Mcp = uc(:,1:N) * uc(:,1:N)' * Mc;figure(1);x0 = [ ];y0 = [ ];x1 = [ ];y1 = [ ];x2 = [ ];y2 = [ ];for i = 1:19v = matricize(Mcp(:,i), 6)';x0 = [x0 v(1,1:5)];y0 = [y0 v(2,1:5)];x1 = [x1 v(3,1:5)];y1 = [y1 v(4,1:5)];x2 = [x2 v(5,1:5)];y2 = [y2 v(6,1:5)];endxmin = min(min([x0(:) x1(:) x2(:)]));xmax = max(max([x0(:) x1(:) x2(:)]));ymin = min(min([y0(:) y1(:) y2(:)]));ymax = max(max([y0(:) y1(:) y2(:)]));axisval = [xmin xmax ymin ymax];for i = 1:95hold offplot([x1(i) x0(i) x2(i)], [y1(i) y0(i) y2(i)] + 30);hold onplot([x1(i) x0(i) x2(i)], [y1(i) y0(i) y2(i)] + 30, `o`);axis([0 100 5 55]);% axis(axisval);pause(0.1);endend%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%if (1)k = 5;s = 10;ndeg = 3;nRows = ndeg *nfr = 100;c = 1;for i = 1:k:(nfr-s+1)v = vectorize(Aa(:, i:i+s-1)');Ma(1:nRows,c) = v;c = c + 1;end[ua, sa, va] = svd(Ma);if (0)figure(4);hold offplot(sqrt(diag(sa)), `--`);hold onplot(sqrt(diag(sc)));endMap = ua(:,1:N) * ua(:,1:N)'* Ma;figure(2);x0 = [ ];y0 = [ ];x1 = [ ];y1 = [ ];x2 = [ ];y2 = [ ];for i = 1:19v = inatricize(Map(:,i), 3)';x0 = [x0 v(1,1:5)];y0 = [y0 v(2,1:5)];x1 = [x1 v(1,1:5) + 5*cos(v(3,1:5))];y1 = [y1 v(2,1:5) + 5*sin(v(3,1:5))];x2 = [x2 v(1,1:5) + 5*cos(v(3,1:5))];y2 = [y2 v(2,1:5) - 5*sin(v(3,1:5))];endif (1)hold offplot([O], [0]);axis([0 50 5 55]);endhold onfor i = 1:95plot([x1(i) x0(i) x2(i)], [y1(i) y0(i) y2(i)] + 30);hold onplot([x1(i) x0(i) x2(i)], [y1(i) y0(i) y2(i)] + 30, `o`);axis([0 100 5 55]);hold offpause(0.1);endendfile: /homes/leventon/Matlab/matricize.mfunction M = inatricize(v, ncol)ncol = round(ncol);nrow = round(length(v) / ncol);M = zeros(nrow, ncol);for i = 1:nrowb = (i-1)*ncol+1;e = i*ncol;M(i,:) = v(b:e)';endfile: /homes/leventon/Matlab/mcylmodel.mfunction person = mcylmodel(A, v)pts = A(2:38, :);utorso{1} = `upper torso`,utorso{2} = norm(pts(29,:)-pts(34,:))/2;utorso{3} = (norm(pts(29,:)-pts(12,:)) + norm(pts(34,:)-pts(16,:)))/2;utorso{4} = norm(pts(1,:)-pts(3,:));utorso{5} = 0;utorso{6} = 0;utorso{7} = 0;utorso{8} = 0;utorso{9} = 0;utorso{10} = [ ];ltorso{1} = `lower torso`;ltorso{2} = norm(pts(29, :)-pts(34,:))/2;ltorso{3} = (norm(pts(29,:)-pts(12,:)) + norm(pts(34,:)-pts(16,:)))/2;ltorso{4} = 0;ltorso{5} = 0;ltorso{6} = 0;ltorso{7} = 0;ltorso{8} = 0;ltorso{9} = 0;ltorso{10} = 0;head{1} = `head`;head{2} = norm(pts(21,:) - pts(4,:))/2;head{3} = norm(pts(21,:) - pts(4,:))/2;head{4} = norm(pts(21,:) - sum(pts([15 11 33 28],:))/4) * 0.7;head{5} = 0;head{6} = 0;head{7} = norm(pts(1,:)-pts(3,:));head{8} = 0;head{9} = 0;head{10} = [ ];leye{1} = `left eye`;leye{2} = head{2}/4;leye{3} = head{2}/4;leye{4} = head{2}/6;leye{5} = head{2}/2;leye{6} = head{2}/sqrt(2);leye{7} = head{4}*0.7;leye{8} = pi/2;leye{9} = 0;leye{10} = 0;leye{11} = [ ];reye{1} = `right eye`;reye{2} = head{2}/4;reye{3} = head{2}/4;reye{4} = head{2}/6;reye{5} = head{2}/2;reye{6} = head{2}/sqrt(2);reye{7} = head{4}*0.7;reye{8} = pi/2;reye{9} = 0;reye{10} = 0;reye{11} = [ ];ruarm{1} = `rt upper arm`;ruarm{2} = (norm(pts(34,:)-pts(16,:)) + norm(pts(35,:)-pts(17,:))) /4;ruarm{3} = ruarm{2};ruarm{4} = (norm(pts(34,:)-pts(35,:)) + norm(pts(16,:)-pts(17,:))) /2;ruarm{5} = utorso{2};ruarm{6} = 0;ruarm{7} = utorso{4}-ruarm{2};ruarm{8} = [ ];ruarm{9} = [ ];ruarm{10} = 0;rlarm{1} = `rt lower arm`;rlarm{2} = (norm(pts(37,:)-pts(18,:)) + norm(pts(35,:)-pts(17,:))) /4;rlarm{3} = rlarm{2};rlarm{4} = (norm(pts(35,:)-pts(37,:)) + norm(pts(17,:)-pts(18,:))) /2;rlarm{5} = 0;rlarm{6} = 0;rlarm{7} = ruarm{4};rlarm{8} = [ ];rlarm{9} = [ ];rlarm{10} = 0;rhand{1} = `rt hand`;rhand{2} = norm(pts(36,:)-pts(37,:))/2 * 0.5;rhand{3} = norm(pts(36,:)-pts(37,:))/2 * 0.5;rhand{4} = norm(pts(36,:)-pts(37,:));rhand{5} = 0;rhand{6} = 0;rhand{7} = rlarm{4};rhand{8} = 0;rhand{9} = 0;rhand{10} = 0;luarm{1} = `it upper arm`;luarm{2} = (norm(pts(29,:)-pts(12,:)) + norm(pts(30,:)-pts(13,:))) /4luarm{3} = luarm{2};luarm{4} = (norm(pts(29,:)-pts(30,:)) + norm(pts(12,:)-pts(13,:))) /2;luarm{5} = -utorso{2};luarm{6} = 0;luarm{7} = utorso{4}-luarm{2};luarm{8} = [ ];luarm{9} = [ ];luarm{10} = 0;llarm{1} = `lt lower arm`;llarm{2} = (norm(pts(32,:)-pts(14,:)) + norm(pts(30,:)-pts(13,:))) /4;llarm{3} = llarm{2};llarm{4} = (norm(pts(30,:)-pts(32,:)) + norm(pts(13,:)-pts(14,:))) /2;llarm{5} = 0;llarm{6} = 0;llarm{7} = luarm{4};llarm{8} = [ ];llarm{9} = [ ];llarm{10} = 0;lhand{1} = `lt hand`;lhand{2} = norm(pts(32,:)--pts 31,:) /2 * 0.5;lhand{3} = norm(pts(32,:)-pts(31,:))/2 * 0.5;lhand{4} = norm(pts(32,:)-pts(31,:));lhand{5} = 0;lhand{6} = 0;lhand{7} = llarm{4};lhand{8} = 0;lhand{9} = 0;lhand{10} = 0;ruleg{1} = `rt upper leg`;ruleg{2} = (norm(pts(1,:)-pts(8,:)) + norm(pts(9,:)-pts(25,:))) /4;ruleg{3} = ruleg{2};ruleg{4} = (norm(pts(1,:)-pts(25,:)) + norm(pts(8,:)-pts(9,:))) /2;ruleg{5} = utorso{2}/2 - ruleg{2};ruleg{6} = 0;ruieg{7} = 0;ruleg{8} = [ ];ruleg{9} = [ ];ruleg{10} = 0;rlleg{1} = `rt lower leg`;rlleg{2} = (norm(pts(26,:)-pts(10,:)) + norm(pts(25,:)-pts(9,:))) /4;rlleg{3} = rlleg{2};rlleg{4} = (norm(pts(25,:)-pts(26,:)) + norm(pts(9,:)-pts(10,:))) /2;rlleg{5} = 0;rlleg{6} = 0;rlleg{7} = ruleg{4};rlleg{8} = [ ];rlleg{9} = [ ];rlleg{10} = 0;rfoot{1} = `rt foot`;rfoot{2} = norm(pts(27,:)-pts(26,:))/5;rfoot{3} = norm(pts(27,:)-pts(26,:))/5;rfoot{4} = norm(pts(27,:)-pts(26,:));rfoot{5} = 0;rfoot{6} = 0;rfoot{7} = rlleg{4};rfoot{8} = [ ];rfoot{9} = [ ];rfoot{10} = 0;luleg{1} = `lt upper leg`;luleg{2} = (norm(pts(1,:)-pts(5,:)) + norm(pts(6,:)-pts(22,:))) /4;luleg{3} = luleg{2};luleg{4} = (norm(pts(1,:)-pts(22,:)) + norm(pts(5,:)-pts(6,:))) /2;luleg{5} = -utorso{2}/2+luleg{2};luleg{6} = 0;luleg{7} = 0;luleg{8} = [ ];luleg{9} = [ ];luleg{10} = 0;llleg{1} = `lt lower leg`;llleg{2} = (norm(pts(23,:)-pts(7,:)) + norm(pts(22,:)-pts(6,:))) /4;llleg{3} = llleg{2};llleg{4} = (norm(pts(22,:)-pts(23,:)) + norm(pts(6,:)-pts(7,:))) /2;llleg{5} = 0;llleg{6} = 0;llleg{7} = luleg{4};llleg{8} = [ ];llleg{9} = [ ];llleg{10} = 0;lfoot{1} = `rt foot`;lfoot{2} = norin(pts(24,:)-pts(23,:))/5;lfoot{3} = norm(pts(24,:)-pts(23, :))/5;lfoot{4} = norm(pts(24, :)-pts(23, :));lfoot{5} = 0;lfoot{6} = 0;lfoot{7} = rlleg{4};lfoot{8} = [ ];lfoot{9} = [ ];lfoot{10} = 0;person = cell(1,8);person{1} = `person`;if (nargin >= 2)person{2} = v(1);person{3} = v(2);person{4} = v(3);person{5} = v(4);person{6} = v(5);person{7} = v(6);utorso{10} = v(7);ruarm{8} = v(8);ruarm{9} = v(9);rlarm{8} = v(10);rlarm{9} = v(11);luarm{8} = v(12);luarm{9} = v(13);llarm{8} = v(14);llarm{9} = v(15);ruleg{8} = v(16);ruleg{9} = v(17);rlleg{8} = v(18);rlleg{9} = v(19);luleg{8} = v(20);luleg{9} = v(21);llleg{8} = v(22);llleg{9} = v(23);rfoot{8} = v(24);rfoot{9} = v(25);lfoot{8} = v(26);lfoot{9} = v(27);head{10} = v(28);endhead{11} = {leye reye};rhand{11} = { };rlarm{11} = {rhand};ruarm{11} = {rlarm};lhand{11} = { };llarm{11} = {lhand};luarm{11} = {llarm};rfoot{11} = { };rlleg{11} = {rfoot};ruleg{11} = {rlleg};lfoot{11} = { };llleg{11} = {lfoot};luleg{11} = {llleg};utorso{11} = {ruarm luarm head};ltorso{11} = {ruleg luleg};% utorso{11} = {ruarm};% ltorso{11} = { };person{8} = {utorso ltorso};file /homes/leventon/Matlab/mdecode.mfunction A = mdecode(M, nmk, s, k)[nrow, ncol] = size(M);nRows = nmk * 3;nfr = (ncol-1)*k+s;MA = zeros(nmk*3*nfr, 1);MC = zeros(nfr, 1);x = linspace(-(s-1)/2, (s-1)/2, s)`;sig = s/2-1;mask = exp(-x.*x/(sig*sig));% mask = ones(s,1);X mask = [ones(s/2,1); 0.000001*ones(s/2,1)];b = 1;for i = 1:ncolMC(b:b+s-1) = MC(b:b+s-1) + mask;for j = 1:sb2 = (b - 1) * 3*nmk + 1;e2 = b2 + nRows -1;MA(b2:e2) = MA(b2:e2) + mask(j) * M((j-1)*nRows+1:j*nRows, i);b = b + 1;endb = b + -s + k;endfor i = 1:nfrb = (i-1)*nmk*3 + 1;e = i*nmk*3;MA(b:e) = MA(b:e) / MC(i);endA = [nfr nmk 30; matricize(MA, 3)];file: /homes/leventon/Matlab/xndecode2.mfunction A = mdecode2(M, nmk, s, k)[nrow, ncol] = size(M);nRows = nmk * 2;nfr (ncol-1)*k+s;MA = zeros(nmk*2*nfr, 1);MC = zeros(nfr, 1);x = linspace(-(s-1)/2, (s-1)/2, s)';sig = s/2-1;mask = exp(-x.*x/(sig*sig));% mask = ones(s,1);b = 1;for i = 1:ncolMC(b:b+s-1) = MC(b:b+s-1) + mask;for j = 1:sb2 = (b - 1) * 2*nmk + 1;e2 = b2 + nRows - 1;MA(b2:e2) = MA(b2:e2) + mask(j) * M((j-1)*nRows+1:j*nRows, i);b = b + 1;endb = b + -s + k;endfor i = 1:nfrb = (i-1)*nml*2+ 1;e = i*nmk*2;MA(b:e) = MA(b:e) / MC(i);endA = [nfr nmk; matricize(MA, 2)];file: /homes/leventon/Matlab/mdecodec.,function A = mdecodec(M, nmk, s, k)[nrow, ncol] = size(M);nRows = nmk * 3;nfr = (ncol-1)*k+s;MA = zero(nml*3*nfr, ncol);MC = zeros(nfr, 1);x = linspace(-(s-1)/2, (s-1)/2, s)';sif = s/2-1;mask = exp(-x.*x/(sig*sig));% mask = ones(s,1);% mask = [ones(s/2,1); 0.000001*ones(s/2,1)];MM = zeros(3, nmk, s, ncol);for i = 1:ncolb = 1;for j = 1:sfor kk = 1:nmkMM(:, kk, j, i) = M(b:b+2, i);b = b + 3;endendendMA = zeros(nmk*3(\*nfr, 1);cols = [ ];for t = 1:nfrrl,3 if (mod(t-1, k) == 0)newcol = (t-1)/k+1;if (newcol <= ncol)if (isempty(cols) == 0)ycom = 0;msum = 0; for i = cols j = t - (i-1)*k; ycom = ycom + mask(j) * sum(MM(2, :, j, i),2)/nmk; msum = msum + mask(j); end ycom = ycom / msum; thisycom = sum(MM(2, :, 1, newcol), 2)/nmk; MM(2, :, :, newcol) = MM(2, :, :, newcol) - thisycom + ycom;endcols = [cols newcol];endendmsum = 0;ycom = 0;cols2 = cols;for i = cols2j = t - (i`1)*k;if (j == s)n = length(cols);cols = cols(2:n);endb = (t-1)*nmk*3 + 1;MA(b:b+nRows-1) = MA(b:b+nRows-1) + mask(j) * vectorize(MM(:, :,j, i)');msum = msum + mask(j);endMA(b:b+nRows-1) = MA(b:b+nRows-1) / msum;endA = [nfr nmk 30; matricize(MA, 3)];file: /homes/leventon/Matlab/mencode.mfunction M = mencode(varargin)params = varargin{nargin};n = nargin - 1;s = params(1);k = params(2);M= [ ];c = 1;A = varargin{1};nmk = A(1,2);ndim = 3;nRows = nmk * s * ndim;for w = 1:nA = varargin{w};nfr = A(1,1);nmkchk = A(1,2);if (nmkchk != nmk)error(`Bad AOA Matrix!`);end[nrowA, ncolA] = size(A);if (ncolA .revreaction.= ndim)error(`Bad AOA Matrix!`);endA = A(2:nrowA, :);ndat = nfr * nmk;for i = 1:k:(nfr-s+1)b = (i-1)*nmk+1;e = b + nmk * s - 1;v = vectorize(A(b:e, :));M(i1nRows,c) = v;c = c + 1;endendfile: /homes/leventon/Matlab/mencode2.mfunction M = mencode2(varargin)params = varargin{nargin};n = nargin - 1;s = params(1);k = params(2);M = [ ];c = 1;A = varargin{1};nmk = A(1,2);nRowS = nmk * s * 2;for w = 1:nA = varargin{w};nfr = A(1,1);nmkchk = A(1,2);if (nmkchk != nmk)error(`Bad ADA Matrix!`);end[nrowA, ncolA] = size(A);if (ncolA != 2)error(`Bad ADA Matrix!`);endA = A(2:nrowA, :);ndat = nfr * nmk;for i = 1:k:(nfr-s+1)b = (i-1)*nmk+1;e = b + nmk * s - 1;v = vectorize(A(b:e, :));M(1:nRows,c) = v;c = c + 1;endendfile: /homes/leventon/Matlab/mimagesc.mfunction mimagesc(I, rlen);[m, n, p] = size(I);if (nargin < 2)py = ceil(sqft(p));elsepy = rlen;endpx = ceil(p/py);O = zeros(m*px, n*py);c = 1;for i=1:m:(m-1)*pxfor j=1:n:(n-1)*py0(i:i+m-1, j:j+n-1) = I(:,:,c);c = c + 1;if (c > p)breakendendif (c > p)breakendendimagesc(O');axis imagefile: /homes/levent/Matlab/mkr2ang.mfunction[allv, putor, pltor] = mkr2and(A)nfr = A(1,1);nmk = A(1,2);allpts = A(2:nfr*nmk+1, :);allv = zeros(28, nfr);for t = 1:nfrpts = allpts((t-1)*nmk+1:t*nmk, :);v= zeros(23, 1);%% Find the x, y, and z vectors of the body (lower torso / hips)z = normalize(((pts(12,:)+pts(16,:)+pts(29,:)+pts(34,:))/4) -pys(1, :));x2 = normalize(pts(8,:)-pts(1,:));x = normalize(x2 - dot(x2, z) * z);y = cross(z, x);v(1:3) = pts(1,:);%%Rlor = [x' y' z'];v(4:6) = rToA(Rltor);%% Now, transform the points into a lower-torso-centered coord%% system.pltor = (Rltor' * pts')';%% Compute the upper torsoz = [0 0 1];x2 = normalize(((pltor(16,:)+pltor(34, :))-(pltor(12, :)+pltor(29, :)))/2);x = normalize(x2 - dot(x2, z) * z);y = cross(z, x);Rutor = [x' y' z'];a = rToA(Rutor);if (abs(a(1)) > pi/2)a = a + pi;endv(7) = a(3);Rutor = aToR([0 0 a(3)]);putor = (Rutor' * pltor`)`;%% Compute the Head Positionz = [0 0 1];y2 = normalize(putor(20,:)-putor(4,:));y = normalize(x2 - dot(x2, z) * z);x = cross(y, z);Rhead = [x' y' z'];a = rToA(Rhead);if (abs(a(i)) > pi/2)a = a + pi;endv(28) = a(3) + pi/2;%% Compute the Right Upper Armz = normalize((putor(17, :)+putor(35, :))/2 -(putor(16, :)+putor(34, :))/2);x2 = normalize(putor(17,:)-putor(35,:));x = normalize(x2 - dot(x2, z) *y = cross(z, x);Rruarm = [x' y' z'];prua = rToA(Rruarm);v(8:9) = a(1:2);Rruarm = aToR([a(1:2) 0]);prua = (Rruarm' * putor`)`;%% Compute the Right Lower Armz = normalize((prua(18,:)+prua(37,:)) - (prua(17,:)+prua(35,:)));x2 = normalize(prua(18,:)-prua(37,:));x = normalize(x2 - dot(x2, z) * z);y = cross(z, x);Rrlarm = [x' y' z'];a = rToA(Rrlarm);v(10:11) = a(1:2);%% Compute Left Upper Armz = normalize((putor(13, :)+putor(30, :))/2 -(putor(12, :)+putor(29, :))/2);x2 = normalize(putor(13, :)-putor(30,:));x = normalize(x2 - dot(x2, z) * z);y = cross(z, x);Rluarm = [x' y' z'];a = rToA(Rluarm);v(12:13) = a(1:2);Rluarm = aToR([a(1:2) 0]);plua = (Rluarm'* putor`)`;%% Compute the Left Lower Armz = normalize((plua(14,:)+plua(32,:)) - (plua(13,:)+plua(30,:)));x2 = normalize(plua(32, :)-plua(14,:));x = normalize(x2 - dot(x2, z) * z);y = cross(z, x);Rllarm = [x' y' z'];a = rToA(Rllarm);v(14:15) = a(1:2);%% Compute Right Upper Legz = normalize((pltor(9,:)+pltor(25,:))12 - (pltor(1,:)+pltor(8,:))/2);x2 = normalize(pltor(9, :)-pltor(25,:));x = normalize(x2 - dot(x2, z) * z);y = cross(z, x);Rrlleg = [x' y' z'];a = rToA(Rrlleg);v(16:17) = a(1:2);Rrlleg = aToR([a(1:2) 0]);prll = (Rrlleg' * pltor`)`;%% Compute the Right Lower Legz = normalize((prul(10,:)+prul(26,:)) - (prul(9,:)+prul(25,:)));x2 = normalize(prul(10,:)-prul(26,:));x = normalize(x2 - dot(x2, z) * z);y = cross(z, x);Rrlleg = [x' y' z'];a = rToA(Rrlleg);v(18:19) = a(1:2);Rrlleg = aToR([a(1:2) 0]);prll = (Rrlleg'* prul`)`;%% Compute the Right Footz = normalize(prll(27,:)-prll(26,:));x2 = normalize(prll(26,:)-prll(10,:));x = normalize(x2 - dot(x2, z) *y = cross(z, x);Rrfoot = [x' y' z'];a = rToA(Rrfoot);v(24:25) = a(1:2);%% Compute Left Lower Legz = normalize((pltor(6,:)+pltor(22,:))/2 - (pltor(1,:)+pltor(5,:))/2);x2 = normalize(pltor(6,:)-pltor(22,:));x = normalize(x2 - dot(x2, z) *y = cross(z, x);Rluleg = [x' y' z'];a = rToA(Rluleg);v(20:21) = a(1:2);Rluleg = aToR([a(1:2) 0]);plul = (Rluleg' * pltor`)`;%% Compute the Left Lower Legz = normalize((plul(7,:)+plul(23,:)) - (plul(6,:)+plul(22,:)));x2 = normalize(plul(23,:)-plul(7,:));x = normalize(x2 - dot(x2, z) *y = cross(z, x);Rllleg = [x' y' z'];a = rToA(Rllleg);v(22:23) = a(1:2);Rllleg = aToR([a(1:2) 0]);plll = (Rllleg' * plul`)`;%% Compute the Left Footz = normalize(plll(24,:)-plll(23,:));x2 = normalize(plll(23,:)-plll(7,:));x = normalize(x2 - dot(x2, z) * z);y = cross(z, x);Rlfoot = [x' y' z'];a = rToA(Rlfoot);v(26:27) = a(1:2);allv(:,t) = v;endfile: /homes/leventon/Matlab/mnbrak.m%% mnbrak.m algo from p. 400 on num rec. in C.function [ax, bx, cx, fa, fb, fc] = mnbrak(ax, bx, FUN);GOLD = 1.618034;GLIMIT = 100.0;TINY = 1.0e-20;evalstr = [FUN];fa = eval([evalstr, `(ax)`]);fb = eval([evalstr, `(bx)`]);if (fb > fa)dum = ax; ax = bx; bx = dum;dum = fa; fa = fb; fb = dum;endcx = bx + GOLD * (bx - ax);fc = eval([evalstr, `(cx)`]);while (fb > fc)r = (bx - ax) * (fb - fc);q = (bx - cx) * (fb - fa);u = bx - ((bx - cx) * q - (bx - ax)*r) / . . .(2.0*abs(max(abs(q-r) ,TINY)) * sign(q-r));ulim = bx + GLIMIT*(cx-bx);if ( (bx-u)*(u-cx) > 0.0)fu = eval([evalstr, `(u)`]);if (fu < fc)ax = bx; bx = u; fa = fb; fb = fu;return;elseif (fu > fb)cx = u; fc = fu;return;endu = cx + GOLD*(cx - bx);fu = eval([evalstr, `(u)`]);elseif ( (cx-u)*(u-ulim) > 0.0)fu = eval([evalstr, `(u)`]);if (fu < fc)bx = cx; cx = u; u = cx + GOLD*(cx-bx);fb = fc; fc = fu; fu = eval([evalstr, `(u)`]);endelseif ( (u-ulim)*(ulim-cx) >= 0.0)u=ulim;fu = eval([evalstr, `(u)`]);endax = bx; bx = cx; cx = u;fa = fb; fb = fc; fc = fu;endfile: /homes/leventon/Matlab/mode.mfunction [m, c] = mode(I, d, bucket)%% MODE Mode computation%%%% m = mode(x) computes the mode of vector x%% m = mode(I) computes the mode along the columns of matrix I%% m = mode(I, dim) computes the mode along dimension dim of I%%if (nargin < 2), d = 1; endif (nargin < 3), bucket = 1; endif (bucket > 1)I = round(I / bucket);endif ((ndims(squeeze(I)) == 1) & (nargin < 2))I = I(:);endnd = ndims(I);if ((d < 1) .vertline. (d > nd))error(`Bad Dimension`)end%% Permute the input so that the dimension we want to work over isfirstp = [d 1:d-1 d+1:nd];I = permute(I, p);if (0)%% This is way too slow . . . (Using HIST)%% Make sure it's not completely uniformflr = floor(min(I(:)));cel = cell(max(I(:)));if (flr == cel)m = I(1);c = 1;returnend[N, X] = hist(I(:,:), flr:cel);[c, which] = max(N);m = X(which);else%% Try this: (believe it or not, it's faster by 10x)%% Idea: sort by columns, and traverse each row looking for%% most occurances of the same numberJ = I(:,:);J = sort(J, 1);[a, b] = size(J);m = zeros(1,b);c = zeros(1,b);for j = 1:bbestc = 0;count = 1;m(j) = J(1, j);for i = 2:arl,7 if (J(i, j) == J(i-1, j)) count = count + 1; if (count > bestc) bestc = count; m(j) = J(i, j); endelse count = 1;endendc(j) = bestc;endend%% Need to reshape it backisize = size(I);newshape = num2cell(isize(2:nd));m = reshape(m, newshape{:}, 1);m = shiftdim(m, -1);m = ipermute(m, p);c = reshape(c, newshape{:}, 1);c = shiftdim(c, -1);c = ipermute(c, p);c = c / size(I(:,:), 1);if (bucket > 1)m = m * bucket + floor(bucket/2);endfile: /homes/leventon/Matlab/multiproj.mfunction P = multiproj(A, U, ns)M = mencode(A, [10 5]);j = 0;[nrow, ncol] = size(A);minx = min(A(2:nrow, 1));maxx = max(A(2:nrow, 1));difx = maxx-minx;for i = nsif (i == 0)Ap = mdecode(M, 37, 10, 5);elseMp = U(:,1:i) * U(:,1:i)' * M;Ap = mdecode(Mp, 37, 10, 5);endj = j + 1;[nrow, ncol] = size(Ap);Ap(2:nrow, 1) = Ap(2:nrow, 1) + (difx * j * 2);P{j} = Ap;endfile: /homes/leventon/Matlab/noisescaoa.mfunction A2 = noisescaoa(A1, s)if (s == 0)A2 = A1;returnendnfr = A1(1,1);n = nfr * 37;A2 = A1(2:n+1, :);for i = 1:ntheta = rand * 2 * pi - pi;v = [cos(theta) sin(theta)] * s * randn(1);A2(1, :) = A2(j, :) + v;endA2 = [[nfr 37]; A2];file: /homes/leventon/Matlab/normalize.mfunction v = normalize(u)nu = norm(u);if (nu == 0)v = u;elsev = u / norm(u);endfile: /homes/leventon/Matlab/normsq.mfunction n = normsq(v)n = sum(v.*v);file: /homes/leventon/Matlab/pad.mfunction I = pad(I, xp, yp, val)I = [val*ones(size(I,1), xp) I val*ones(size(I,1), xp)];I = [val*ones(yp, size(I,2)); I; val*ones(yp, size(I,2))];file: /homes/leventon/Matlab/pnmRead.m% IM = pnmRead( FILENAME )%% Load a pgm or ppm image into a MatLab matrix.% This format is accessible from the XV image browsing utility.% Works for 8bit gray images (raw or ascii)% and 24bit rgb images.% Hany Farid, Spring '96. Modified by Eero Simoricelli, 6/96.% Modified Leventon '98function im = ppmRead( fname );[fid,msg] = fopen( fname, `r`);if (fid == -1)error(msg);end%%% First line contains ID string:%%% "P1" = ascii bitmap, "P" = ascii greymap,%%% "P3" = ascii pixmap, "P4" = raw bitmap,%%% "P5" = raw greymap, "P6" = raw pixmapTheLine = fgetl(fid);format = TheLine;if (format(i) .sup..about. = `P`)error(`Not a PNM image`);end%%% Any number of comment linesTheLine = = fgetl(fid);while (isempty(TheLine) .vertline. (TheLine(1) == `#`))TheLine = fgetl(fid);end%%% dimensionsszv = sscanf(TheLine,`%d`,3);xdim = szv(1);ydim = szv(2);sz = xdim * ydim;if (length(szv) == 3)maxcal = szv(3);else%%% Maximum pixel valueTheLine = fgetl(fid);maxval = sscanf(TheLine, `%d`,1);end%%im = zeros(dim,1);if ((format(2) == `2`) .vertline. (format(2) == `5`))if (format(2) == `2`)[im,count] = fscanf(fid, `%d`,sz);else[im,count] = fread(fid,sz, `uchar`);endfclose(fid);if (count == sz)im = reshape( im, xdim, ydim )';elsefprintf(1, `Warning: File ended early!`);im = reshape( [im ; zeros(sz-count,1)], xdim, ydim)`;endelseif ((format(2) == `3`) .vertline. (format(2) == `6`))sz = sz*3;if (format(2) == `3`)[im,count] = fscanf(fid,`%d`,sz);else[im,count] = fread(fid,sz,`uchar`);endfclose(fid);if (count == sz)im = permute(reshape( im, 3, xdim, ydim), [3 2 1]);elsefprintf(1, `Warning: File ended early!`);im = perinute(reshape( [im ; zeros(sz-count,1)], 3, xdim, ydim), [32 1]);endendfile: /homes/ieventon/Matlab/ppmWrite.m% RANGE = ppmWrite(MTX, FILENAME, RANGE, TYPE)% Write a MatLab matrix to a pgm (graylevel image) file.% This format is accessible from the XV image browsing utility.% RANGE (optional) is a 2-vector specifying the values that map to% black and white, respectiveiy. Passing a value of `auto` (default)% sets RANGE=[min,max] (as in MatLab's imagesc). `auto2` sets% RANGE=[mean-2*stdev, mean+2*stdev]. `auto3` sets% RANGE=[p1-(p2-p1)/8, p2+(p2-p1)/8], where p1 is the 10th percentile% value of the sorted MATRIX samples, and p2 is the 90th percentile% value.%% TYPE (optional) should be `raw` or `ascii`. Defaults to `raw`.% Hany Farid, Spring '96. Modified by Eero Simoncelli, 6/96.function range = pgrnWrite(mtx, fname, range, type );[fid,msg] = fopen( fname, `w` );if (fid == -1)error(msg);end%-----------------------------------------------------------------file: %% optional ARGS:if (exist(`range`) .sup..about. = 1)range = `auto`;endif (exist(`type`) .sup..about. = 1)type = `raw`;endfile:%% Automatic range calculation:if (strcmp(range,`auto1`) .vertline. strcmp(range,`auto`))[mn,mx] = range2(mtx);range = [mn,mx];elseif strcmp(range,`auto2`)stdev = sqrt(var2(mtx));av = mean2(mtx);range = [av-2*stdev,av+2*stdev]; % MAGIC NUMBER: 2 stdevselseif strcmp(range, `auto3`)percentile = 0.1; % MAGIC NUMBER: 0<p<0.5[N,X] = histo(mtx);binsz = X(2)-X(1);N = N+1e-10; % Ensure cumsum will be monotonic for call to interplcumN = [0, cumsum(N)]/sum(N);cumX = [X(1)-binsz, X] + (binsz/2);ctrRange = interp1(cumN,cumX, [percentile, 1-percentile]);range = mean(ctrRange) + (ctrRange-mean(ctrRange))/(1-2*percentile);elseif isstr(range)error(sprintf(`Bad RANGE argument: %s`,range))endif ((range(2) - range(1)) <= eps)range(1) = range(1) - 0.5;range(2) = range(2) + 0.5;end%%% First line contains ID string:%%% "P1" = ascii bitmap, "P2" = ascii greymap,%%% "P3" = ascii pixmap, "P4" = raw bitmap,%%% "P5" = raw greymap, "P6" = raw pixmapif strcmp(type, `raw`)fprintf(fid, `P6.backslash.n`);format = 6;elseif strcmp(type, `ascii`)fprintf(fid, `P3.backslash.n`);format = 3;elseerror(sprintf(`PGMWRITE: Bad type argument: %s`,type));endfprintf(fid,`# MatLab PGMWRITE file, saved %s.backslash.n`,date);%%% dimensionsfprintf(fid,`%d %d.backslash.n`,size(mtx,2),size(mtx,1));%%% Maximum pixel valuefprintf(fid, `255.backslash.n`);%% MatLab's "fprintf" floors when writing floats, so we compute%% (mtx-r1)*255/(r2-r1)+0.5mult = (255 / (range(2)-range(1)));mtx = (mult * mtx) + (0.5 - mult * range(1));mtx = max(-0.5+eps,min(255.5-eps,mtx));mtx = reshape(permute(mtx, [3 2 1]), 3*size(mtx,2), size(mtx,1));if (format == 3)count = fprintf(fid,`%d `,mtx);elseif (format == 6)count = fwrite(fid,mtx,`uchar`);endfclose(fid);if (count .sup..about. = size(mtx,1)*size(mtx,2))fprintf(1,`Warning: File output terminated early!`);end%%% TEST:% foo = 257*rand(100)-1;% pgmWrite(foo,`foo.pgm`,[0 255]);% f002=pgmRead(`foo.pgm`);% size(find((foo2-round(foo)).sup..about. =0))% foo(find((foo2-round(foo)).sup..about. =0))file: /homes/leventon/Matlab/proj2d.mfunction A2 = proj2d(A3)n = size(A3, 1);A3a = A3(2:n, :);A2a = [A3a(:,1) A3a(:,3)];A2 = [A3(1, 1:2); A2a];file: /homes/leventon/Matlab/project.mfunction [Ap, v, Mp, M] = project(A, U, n)M = mencode(A, [10 5]);v = U(:,1:n)' * M;Mp = U(:,1:n) * v;Ap = mdecode(Mp, 37, 10, 5);file: /homes/leventon/Matlab/project2.mfunction [Ap, v, Mp, M] = project2(A, U, n)M = mencode2(A, [10 5]);v = U(:,1:n)' * M;Mp = U(:,1:n) * v;Ap = mdecode2(Mp, 37, 10, 5);file: /homes/leventon/Matlab/project23.mfunction [A3p, A2p, Mp3] = project23(A2, M3, U, S, V, n)M2 = mencode2(A2, [10 5]);nEx = size(V, 1);a = U(:,1:n)' * M2;Si = [(diag(1./(diag(S(:,1:n))))); zeros(nEx-n, n)];Mp2 = U(:, 1:n) * a;A2p = mdecode2(Mp2, 37, 10, 5);Mp3 = M3 * V * Si * a;A3p = mdecodec(Mp3, 37, 10, 5);file: /homes/leventon/Matlab/project23a.mfunction [A3p, A2p, Q] = project23a(A2, M3, U, S, V, nEv, lambda)M2 = mencode2(A2, [10 5]);[m, n] = size(M3);nmk = 37;s = m / nmk / 3;k = nmk * 3 * (s-2);nEx = size(V, 1);uP = U(:, 1 :nEv);a = Up' * M2;Si = [(diag(1./(diag(S(:,1:nEv))))); zeros(nEx-nEv, nEv)];if (0)CM = zeros(3*s, m);mtmp1 = vectorize([ones(1,nmk)/nmk; zeros(2, nmk)]`)`;mtmp2 = vectorize([zeros(1,nmk); ones(1,nmk)/nmk; zeros(1, nmk)]`)`;mtmp3 = vectorize([zeros(2, nmk); ones(1,nmk)/nmk]`)`;for i=1:sCM((i-1)*3+1,:) = [zeros(1, (i-1)*xnmk*3) mtmp1 zeros(1,(s-i)*nmk*3)];CM((i-1)*3+2,:) = [zeros(1, (i-1)*nmk*3) mtmp2 zeros(1,(s-i)*nmk*3)];CM((i-1)*3+3,:) = [zeros(1, (i-1)*nmk*3) mtmp3 zeros(1,(s-i)*nmk*3)];ende = ones(k,1);D = spdiags([-e/2 e/2], [0 6], k, s*3);A = spdiags([e -2*e e], [0 3 6], k, s*3);M4 = M3 * V * Si;Q = M4'*CM'*A'*D*CM*M4;elsee = ones(k,1);D = spdiags([-e/2 e/2], [0 nmk*6], k, m);A = spdiags([e -2*e e], [0 nmk*3 nmk*6], k, m);M4 = M3 * V * Si;Q = M4'*A'*D*M4;endaw = inv(Up'*Up + lambda * Q) * Up' * M2;Mp2 = Up * aw;A2p = mdecode2(Mp2, 37, 10, 5);Mp3 = M3 * V * Si * aw;A3p = mdecode(Mp3, 37, 10, 5);file: /homes/leventon/Matlab/project23d.mfunction [A3p, A2p, Q] = project23d(A2, M3, U, S, V, nEv, lambda)M2 = mencode2(A2, [10 5]);[m, n] = size(M3);nmk = 37;s = m / nmk / 3;k = nmk * 3 * (s-2);nEx = size(V, i);Up = U(:, i:nEv);a = Up' * M2;Si = [(diag(1./(diag(S(:,1:nEv))))); zeros(nEx-nEv, nEv)];if (0)CM = zeros(3*s, m);mtmp1 = vectorize([ones(1,nmk)/nmk; zeros(2, nmk)] `)`;mtmp2 = vectorize([zeros(1,nmk); ones(1,nmk)/nmk; zeros(1, nmk)`)`;mtmp3 = vectorize([zeros(2, nmk); ones(1,nmk)/nmk]`)`;for i=1:sCM((i-1)*3+1,:) = [zeros(1, (i-1)*nmk*3) mtmp1 zeros(1,(s-i)*nmk*3)];CM((i-1)*3+2,:) = [zeros(1, (i-1)*nmk*3) mtmp2 zeros(1,(s-i)*nmk*3)];CM((i-1)*3+3,:) = [zeros(1, (i-1)*nmk*3) mtmp3 zeros(1,(s-i)*nmk*3)];ende = ones(k,1);D = spdiags([-e/2 e/2], [0 6], k, s*3);A = spdiags([e -2*e e], [0 3 6], k, s*3);M4 = M3 * V * Si;Q = M4'*CM'*A'*D*CM*M4;elsee = ones(k,1);D = spdiags([-e/2 e/2], [0 nmk*6], k, m);A = spdiags([e -2*e e], [0 nmk*3 nmk*6], k, m);M4 = M3 * V * Si;Q = M4'*D'*D*M4;endaw = inv(Up'*Up + lambda * Q) * Up' * M2;Mp2 = Up * aw;A2p = mdecode2(Mp2, 37, 10, 5);Mp3 = M3 * V * Si * aw;A3p = mdecodec(Mp3, 37, 10, 5);file: /homes/leventon/Matlab/project23dy.mfunction [A3p, A2p, Q] = project23dy(A2, M3, U, S, V, nEv, lambda)M2 = mencode2(A2, [10 5]);]m, n] = size(M3);nmk = 37;s = m / nmk / 3;k = nmk * 3 * (s-2);nEx = size(V, 1);Up = U(:, 1:nEv);a = Up' * M2;Si = (diag(1./(diag(S(:,1:nEv))))); zeros(nEx-nEv, nEv)];if (0)CM = zeros(3*s, m);mtmp1 = vectorize([ones(1,nmk)/nmk; zeros(2, nmk)]`)`;mtmp2 = vectorize([zeros(1,nmk); ones(1,nmk)/nmk; zeros(1, nmk)]`)`;mtmp3 = vectorize([zeros(2, nmk); ones(1,nmk)/nmk]`)`;for i=1:sCM((i-1)*3+1,:) = [zeros(1, (i-1)*nmk*3) mtmp1 zeros(1,(s-i)*nmk*3)]CM((i-1)*3+2,:) = [zeros(1, (i-1)*nmk*3) mtmp2 zeros(1,(s-i)*nmk*3)]CM((i-1)*3+3,:) = [zeros(1, (i-1)*nmk*3) mtmp3 zeros(1,(s-i)*nmk*3)];endvectorize([zeros(k/3,1) ones(k/3,1) zeros(k/3)]);D 342 spdiags([-e/2 e/2], [0 6], k, s*3);M4 = M3 * V * Si;Q = M4'*C'*D'*D*CM*M4;elsee = vectorize([zeros(k/3,1) ones(k/3,1) zeros(k/3)]);D = spdiags([-e/2 e/2], [0 nmk*6], k, m);M4 = M3 * V * Si;Q = M4'*D'*D*M4;endaw = inv(Up'*Up + lambda * Q) * Up' * M2;Mp = Up * aw;A2p = mdecode2(Mp2, 37, 10, 5);Mp3 = M3 * V * Si * aw;A3p = mdecodec(Mp3, 37, 10, 5);file: /homes/leventon/Matlab/project23w.mfunction [A3p, A2p, Q] = project23w(A2, M3, U, S, V, nEv, lambda)M2 = mencode2(A2, [10 5]);[m, n] = size(M3);nmk = 37;s = m / nmk / 3;k = nmk * 3 * (s-2);nEx = size(V, 1);Up = U(:, 1:nEv);a = Up' * M2;Si = [(diag(1./(diag(S(:,1:nEv))))); zeros(nEx-nEv, nEv)];if (0)CM = zeros(3*s, m);mtmp1 = vectorize([ones(1,nmk)/nmk; zeros(2, nmk)]`)`;mtmp2 = vectorize([zeros(1,nmk); ones(1,nmk)/nmk; zeros(1, nmk)]`)`;mtmp3 = vectorize([zeros(2,nmk); ones(1,nmk)/nmk]`)`;for i=1:sCM((i-1)*3+1,:) = [zeros(1, (i-1)*nmk*3) mtmp1 zeros(1,(s-1)*nmk*3)];CM((i-1)*3+2,:) = [zeros(1, (i-1)*nmk*3) mtmp2 zeros(1,(s-i)*nmk*3)];CM((i-1)*3+3,:) = [zeros(1, (i-1)*nmk*3) mtmp3 zeros(1,(s-i)*nmk*3)];ende = ones(k,1);D = spdiags([-e/2 e/2],[0 6], k, s*3);A = spdiags([e -2*e e], [0 3 6], k, s*3);M4 = M3 * V * Si;Q = M4'*CM'*A'*D*CM*M4;elsee = ones(k,1);D = spdiags([-e/2 e/2], [0 nmk*6], k, m);A = spdiags([e -2*e e], [0 nmk*3 nmk*6], k, m);M4 = M3 * V * Si;Q = M4'*A'*D*M4;endaw = inv(Up'*Up + lamLbda * Q) * Up' * M2;Mp2 = Up * aw;A2p = mdecode2(Mp2, 37, 10, 5);Mp3 = M3 * V * Si * aw;A3p = mdecode(Mp3, 37, 10, 5);file: /homes/leventon/Matlabproject23wM.mfunction [A3p, A2p, Q] = project23wM(A2, M3, U, S, V, nEv, lambda)%% Handles multiple gammas . . .M2 = mencode2(A2, [10 5]);[m, n] = size(M3);nmk = 37;s = m / nmk / 3;k = nmk * 3 * (s-2);nEx = size(V, 1);Up = U(:, i:nEv);a = Up' * M2;Si = [(diag(1./(diag(S(:,i:nEv))))); zeros(nEx-nEv, nEv)];if (0)CM = zeros(3*s, m);mtmp1 = vectorize([ones(1,nmk)/nmk; zeros(2, nmk)]`)`;mtmp2 = vectorize([zeros(1,nmk); ones(1,nmk)/nmk; zeros(1, nmk)]`)`;mtmp3 = vectorize([zeros(2, nmk); ones(1,nmk)/nmk]`)`;for i=1:sCM((i-1)*3+1,:) = [zeros(1, (i-1)*nmk*3) mtmp1 zeros(1,(s-i)*nmk*3)];CM((i-1)*3+2,:) = [zeros(1, (i-1)*nmk*3) mtmp2 zeros(1,(s-i)*nmk*3)];CM((i-1)*3+3,:) = [zeros(1, (i-1)*nmk*3) mtmp3 zeros(1,(s-i)*nmk*3)];ende = ones(k,1);D = spdiags([-e/2 e/2], [0 6], k, s*3);A = spdiags([-e 2*e -e], [0 3 6], k, s*3);M4 = M3 * V * Si;Q = M4'*CM'*A'*D*CM*M4;elsee = ones(k,1);D = spdiags([-e/2 e/2], [0 nmk*6], k, m);A = spdiags([-e 2*e -e], [0 nmk*3 nmk*6], k, m);M4 = M3 * V * Si;Q = M4'*A'*D*M4;endaw = inv(Up'*Up + lambda * Q) * Up' * M2;Mp2 = Up * aw;A2p = mdecode2(Mp2, 37, 10, 5);Mp3 = M3 * V * Si * aw;A3p = mdecode(Mp3, 37, 10, 5);file: /homes/leventon/Matlab/projectV.mfunction A3p = projectV(VObs, MV, M3, n)[U, S, V] = svd(MV);% Mo = encodeVis(VObs, 10, 5);Mo = VObs;nEx = size(V,1);a = U(:,1:n)' * Mo;Si = [(diag(1./(diag(S(:,1:n))))); zeros(nEx-n, n)];% Mop = U(:, i:n) * a;% A2p = mdecode2(Mp2, 37, 10, 5);M3p = M3 * V * Si *A3p = mdecodec(M3p, 37, 10, 5);file: /homes/leventon/Matlab/project.sub.-- matrix.mfunction P = project.sub.-- matrix(npts)P = speye(npts*3, npts*3);P = P(sort([1:3:npts*3 3:3:npts*3]), :);file: /homes/leventon/Matlab/pulseify.mfunction [out] = pulseify(seq, pulses.sub.-- per.sub.-- segment, dnsamp)% I = pulseify(seq, pulses.sub.-- per.sub.-- segment, dnsamp)% Converts to a pulse representation for the sequence.% Each line segment of the figure is given pulses.sub.-- per.sub.--segment% point emitters of orientation energy (which will then be sensed% by sensors). The idea is to make a representation which is fast% to render into sensor responses, and easy to get the derivative of% sensor responses with respect to body pose parameters.%% pulses.sub.-- per.sub.-- segment is how many of these point sourcesthere are on% each line segment.% pt.sub.-- pos is a n.sub.-- segments*pulses.sub.-- per.sub.-- segment*number.sub.-- of.sub.-- temporal.sub.-- frames% by 2 (x,y)% array of the positions of each pulse. (Stack pts from% different times on top of each other. To be consistent with% how% marker data is stored in other programs.% pt.sub.-- ori same dimensions, but giving the orientation of eachpulse.%% output: out is a cell array of pt.sub.-- pos and pt.sub.-- ori overdifferent% times:% {pt.sub.-- pos.sub.-- t1}, {pt.sub.-- ori.sub.-- t1},% {pt.sub.-- pos.sub.-- t2}, {pt.sub.-- ori.sub.-- t2},% {pt.sub.-- pos.sub.-- t3}, {pt.sub.-- ori.sub.-- t3},%%% Convert pt.sub.-- ori values to angles consistantly with MEL bytreating% pt.sub.-- ori as "np" in the following, from stickify.m:% thetas = ((np * [cos(oris); sin(oris)])`)`;% val = cos(2*thetas). 2;% I{t} = renderLine2(I{t}, pi(:,[2 1]), p2(:,[2 1]), val);%% A note on coordinate systems: to overlay pulseify outputs on vtk% images, you need to pump the data through worldtocam. that will% invert it (to accomodate matlab image array format annoying position%% of the origin) and scale it properly so that the data could be% overlaid.% For now: give the output in camera coordinates.% To change that to world coordinates, you'd change the lines%% p1 = dc((t-1)*nmk + c(:,1) + 1, :);%% p2 = dc((t-1)*nmk + c(:,2) + 1, :);%% to%% p1 = d((t-1)*nmk + c(:,1) + 1, :);%% p2 = d((t-1)*nmk + c(:,2) + 1, :);% 9/29/97 wtf wrote it, based on stickify.m% 10/10/97 wtf converted output to cell array, rather than% stacked numeric array% 10/15/97 wtf removed reference to oris and Noris.if (nargin < 3)dnsamp = 0;endglobal seqINFOif (issequence(seq))%% It's a sequenceswitch seq.typecase `3d`seq = seqConvert(seq, `2d`);case `2d`otherwiseerror(`Only can pulseify 2d or 3d seqences`);endn = size(seq.data, 1);d = seq.data(2:n, :);N = seq.nframes;nmk = seqINFO.nummkr(seq.subtype);% data in the camera coordinates.dc = worldTocam(d);elseif (size(seq, 2) == 1)%% It's a snippetN = 10;nmk = 37;d = reshape(seq, 3, nmk * N)';% data in the camera coordinates.dc = worldToCam(d) + ones(nmk*N, 1) * ([40 40] - worldToCam([0 00]));end% size of p1 and p2. Initialize this out of the loop, to help savememory (?)% Introduce variable nsegments, which gives the size of p1 and p2.c = seqINFO.connections{1};p1 = zeros(size(c)); p2 = zeros(size(c)); tmp.sub.-- pt= zeros(size(c));nSegments = size(c, 1);out = cell(N,2);for t = i:N% n.sub.-- segments*pulses.sub.-- per.sub.-- segment by 2, for (x,y)pt.sub.-- pos = zeros(nsegments*pulses.sub.-- per.sub.-- segment, 2);pt.sub.-- ori = zeros(nsegments*pulses.sub.-- per.sub.-- segment, 2);% Return output in camera coordinates. See note at top of file.p1 = dc((t-1)*nmk + c(:,1) + 1, :);p2 = dc((t-1)*nmk + c(:,2) + 1, :);p1 = p1 / (2 dnsamp);p2 = p2 / (2 dnsamp);% p1 and p2 are the end points for each segment. Put desired% number of pulses on each. Stack them.for pulse = 1:pulses.sub.-- per.sub.-- segment pt.sub.-- pos((pulse-i)*nSegments+(1:nSegments), :) . . . = p1 + pulse*(p2 - p1)/(pulses.sub.-- per.sub.-- segment+1);end% normalize the orientation vector to unit length, and replicate% for all pulse points on this line segment (same orientation).tmp.sub.-- pt = p2 - p1;tmp.sub.-- pt = tmp.sub.-- pt ./ repmat(sqrt(sum(tmp.sub.-- pt. 2,2)),1,2);pt.sub.-- ori(1:nSegments*pulses.sub.-- per.sub.-- segment, :) . . . = repmat(tmp.sub.-- pt,pulses.sub.-- per.sub.-- segment, 1);out(t,1) = {pt.sub.-- pos};out(t,2) = {pt.sub.-- ori};endfile: /homes/leventon/Matlab/rToA.mfunction [ang, Rx, Ry, Rz] = rToA(r0, flip)r = r0([3 2 1], :)r = r(:,[3 2 1]);if (nargin < 2)flip = 0;endflip = pi * flip;ang = [0 0 0];e = 1e-6;a = atan2e(r(2,1), r(1,1), e) + flip;c = cos(a);s = sin(a);ang(3) = -a;Rz = [c s 0; -s c 0; 0 0 1];Rxy = Rz * r;if (norm([Rxy([1 3], 2)]) < e)c = Rxy(1,1);s = Rxy(3,1);a = atan2e(s, c, e);elsea = atan2e(-Rxy(1,2),Rxy(3,2), e);c = cos(a);s = sin(a);endang(2) = -a;Ry = [c 0 s; 0 1 0; -s 0 c];Rx = (Ry * Rxy)';if (Rx(1,1) < 0)[ang, Rx, Ry, Rz] = rToA(r0, 1);elseang(1) = atan2e(Rx(3,2), Rx(2,2), e);ang = ang([3 2 1]);end% for i=1:3% while (ang(i) < -pi)% ang(i) = ang(i) + pi;% end% while (ang(i) > pi)% ang(i) = ang(i) - pi;% end%endfile: /homes/leventon/Matlab/randTrans.mfunction T = randTrans( )t = rand(3, 1) * 10;theta = rand * 2 * pi;phi = rand * 2 * pi;alpha = rand * 2 * pi;R1 = [cos(theta) sin(theta) 0; -sin(theta) cos(theta) 0; 0 0 1];R2 = [cos(phi) 0 sin(phi); 0 1 0; -sin(phi) 0 cos(phi)];R3 = [1 0 0; 0 cos(alpha) sin(alpha); 0 -sin(alpha) cos(alpha)];R = R1 * R2 * R3;T = [R t; 0 0 0 1];file: /homes/ieventon/Matlab/rbf.m%% rbf.m radial basis function%% Aug. 1, 1997 freeman created.%% alpha is matrix of alpha coefficients.%% alpha has # training examples columns, and%% the function output dimension number of rows.%% x is input vector.%% x.sub.-- train is a matrix of training input vectors.%% x.sub.-- train is of dimension [input dimensionality, #training%% examples].%% sigma is stiffness parameter.function [y, v, sq.sub.-- dist = rbf(x, x.sub.-- train, alpha, sigma);%% Make a matrix with each column the column vector x,%% and subtract that from the training data.%% Then sum squares, to form a vector of distances from the inputs.sq.sub.-- dist = sum((x.sub.-- train - x * ones(1, size(x.sub.-- train,2)) ). 2);v = (exp(- sq.sub.-- dist / (2 * sigma) 2))';y = alpha * v;file: /homes/leventon/Matlab/regularize.mfunction [xall, xsall] = regularize(encobs, MtMs, coeffs, nEv,lambdas, niter, core, clip, sum1)niter = max(niter, 1);if ((core == 0) & (clip == 0) & (sum1 == 0.3))xall = coeffs;returnendnSnip = size(coeffs, 2);size(MtMs);lambda1 = sum1;%% For each snippet in the observationMtM = MtMs;for snip = 1:nSnipif (iscell(MtMs))MtM = MtMs{snip};endREGULARIZING.sub.-- SNIPPET = snip% Select out the initial guess of coefficients for this snippet% (computed by SVD)x0 = coeffs(:,snip);% Select out the observation for this snippetMty = encobs(:,snip);clear xs;lct = 1;x = x0;for lambda = lambdas%% MtMpL1pLIi = pinv(MtM + lambda1 + lambda*eye(size(MtM)));%% Change: instead of using pinv, let's compute this with svd . . . [U, S, V] = svd(MtM + lambda1 + lambda*eye(size(MtM))); Si = spdiags(1./diag(S), 0, size(V,1), nEv); MtMpL1pLIi = V * (Si * U(:,i:nEv)');for ct = 1:niter%%% Step 2. Core / Clip the coeffs% Parameters of the amount of coring . . .ex = 2;pt = max(x)*3/4;% Computing the coring of the coefficientsax = abs(x);xc = x;if (core) xc = ((ax >= pt) .* x) + (sign(x) .* (ax < pt) .*(ax) . ex/pt (ex-1));endif (clip) xc = xc .* (xc > 0);end%%% Step 3. Minimize%% E = (data - [training set] * coeffs) . 2 +%% lambda * (coeffs - clipped.sub.-- coeffs) . 2x = MtMpL1pLIi * (Mty + lambda*xc + lambda1);xs{lct, ct} = x;endlct = lct + 1;endxall(:,snip) = x;xsall{snip} = xs;endfile: /homes/leventon/Matlab/render.mfunction [im, zbuf] = render(cyl, camera, im, zbuf)%% Default Image Size is 256.times.256if (nargin < 4)im = [ ];zbuf = [ ];endif (isempty(im))im = zeros(256);endif (isempty(zbuf))zbuf = ones(256) * inf;endfile: /homes/leventon/Matlab/renderGuy.m%% . . ./Vtk/renderguy.tcl Temp/person.mdl Temp/tmp.seqTemp/tmp-%03d.ppm%% 90 1 1rc = seqINFO.defaults.rendercmd;m = seqINF0.defaults.model;s = seqINFO.defaults.seq;f = seqINFO.defaults.framespec;model = seqINFO.repAngle.model;display = [`-display ` seqINFO.defaults.display];cmd = sprintf(`%s %s %s %s %S 0 &`, rc, display, m, s, f);disp(`renderGuy: Starting Tcl Prograiu (in background)`);unix(cmd);file: /homes/leventon/Matlab/renderLine.mfunction [I, errs] = renderLine(I, pt1, pt2, val, combineType)if (nargin < 5)combineType = `max`;enderrs = 0;x1 = pt1(1);x2 = pt2(1);y1 = pt1(2);y2 = pt2(2);%% Iterate over axis with greater rangedelta = 1.0;dx = x2 - x1;dy = y2 - y1;if ((dx == 0.0) & (dy == 0.0))[I, errs] = renderpoint(I, x1, y1, val, combineType);elseif (abs(dx) > abs(dy))%% iterate over xm = dy/dx;%% make sure x1 <= x2if (x1 > x2)tmp = x1; x1 = x2; x2 = tmp;tmp = y1; y1 = y2; y2 = tmp;end%% plot each pointfor i = x1:delta:x2[I, err] = renderPoint(I, i, y1 + (i - x1)*m, val, combineType);errs = errs + err;endelse%% iter over ym = dx/dy;%% make sure y1 <= y2if (y1 > y2)tmp = x1; x1 = x2; x2 = tmp;tmp = y1; y1 = y2; y2 = tmp;end%% plot each pointfor i = y1:delta:y2]I, err = renderpoint(I, x1 + (i - y1)*m, i, val, combineType);errs = errs + err;endendfile: /homes/leventon/Matlab/renderLine2.mfunction [I, errs] = renderLine(I, pt1, pt2, val, combineType)if (nargin < 5)combineType = `max`;enderrs = 0;%% Iterate over axis with greater rangedelta = 1.0;dp = pt2 - pt1;N = size(pt1, 1);X = [ ];Y = [ ];V = [ ];adp = abs(dp);for t = 1:Nif (dp(t, :) == 0)X = [X; pt1(t,1)];Y = [Y; pt1(t,2)];V = [V; val(t, :)];elseif (adp(t, 1) > adp(t, 2))%% iterate over xm = dp(t, 2)/dp(t, 1);%% make sure x1 <= x2xs = (pt1(t,1):delta:pt2(t,1))';if (isempty(xs))xs = (pt2(t,1):delta:pt1(t,1))';endX = [X; xs];Y = [Y; pt1(t,2) + (xs - pt1(t,1))*m];V = [V; ones(length(xs), 1) * val(t, :)];else%% iterate over ym = dp(t, 1)/dp(t, 2);%% make sure y1 <= y2ys = (pt1(t,2):delta:pt2(t,2))';if (isempty(ys))ys = (pt2(t,2):delta:pt1(t,2))';endY = [Y; ys];X = [X; pti(t,1) + (ys - pt1(t,2))*m];V = [V; ones(length(ys), 1) * val(t, :)];endendX = round(X);Y = round(Y);indx = find( (X>0) & (Y>0) & (X<size(I,1)) & (Y<size(I,2)) );X = X(indx);Y = Y(indx);V = V(indx, :);N = length(X);for t = 1:NI(X(t), Y(t), :) = max(I(X(t), Y(t), :), shiftdim(V(t, :), -1));endfile: /homes/leventon/Matlab/renderPoint .mfunction [I, err] = renderPoint(I, x, y, val, combineType)x = round(x);y = round(y);if ((x < i) .vertline. (y < 1) .vertline. (x > size(I,1)) .vertline. (y >size(I, 2)))err = 1;elseerr = 0;switch combineTypecase `clobber`I(x, y) = val;case `max`I(x, y) = max(I(x, y), val);case `add`I(x, y) = I(x, y) + val;endendfile: /homes/leventon/Matlab/rendercyl.mfunction [im, zbuf] = rendercyl(cyl, color, camera, light, im, zbuf)%% Default Image Size is 256.times.256if (nargin < 5)im = zeros(256);zbuf = ones(256) * inf;endp = cyl(:,1);R = aToR(cyl(:,2)');r = cyl(:,3)';N = 50;T = linspace(0, 2*pi, N);P0 =P0 = R*[r(1) * cos(T); r(2) * sin(T); zeros(1,N)] + (p * ones(1, N));minv = min(P0);maxv = max(P0);minv = max([minv; 1 1 0]);maxv = min([maxv; 256 256 inf]);for y = minv(2):maxv(2)for x = minv(1):maxv(1)endendfile: /homes/leventon/Matlab/renderdisc.mfunction [im, zbuf] = rendercyl(cyl, color, camera, light, im, zbuf)%% Default Image Size is 256.times.256if (nargin < 5)im = zeros(256);zbuf = ones(256) * inf;endp = cyl(:,1);R = aToR(cyl(:,2)');r = cyl(:,3)';N = 50;T = linspace(0, 2*pi, N);P0 = [r(1) * cos(T); r(2) * sin(T); zeros(1,N)];P0 = R*P0 + (p * ones(1, N));minv = min(P0);maxv = max(P0);minv = max([minv; 1 1 0]);maxv = min([maxv; 256 256 inf]);file: /homes/leventon/Matlab/rgbimagesc.mfunction J0 = rgbimagesc(I, t);if (nargin < 2)t = 1.endif (iscell(I))I = I{t};t = 1;end[m, n, p, q] = size(I);I = I(:,:,:,t);J = zeros(m, n, 3);which = 0;for i = 1:pw = ((which == 3) + (which .sup..about. = 3) * which) + 1;J (:,:,w) = max (J (:,:,w), I(:,:,i));which = mod(which + 1, 4);endfor i = 1:3J(:,:,i) = J(:,:,i) + min(min(J(:,:,i)));val = max(max(J(:,:,i)));if (val > eps)J(:,:,i) = J(:,:,i) / val;endend% J = transpose2(J);if (nargout == 1)J0 = J;elseimagesc(J);axis imageendfile: /homes/leventon/Matlab/rotaoa.mfunction B = rotaoa(A, theta)n = size(A, 1);A1 = A(2:n, :);B1 = [(cos(theta)*A1(:,i) + sin(theta)*A1(:,2)) (-sin(theta)*A1(:,1)+ cos(theta)*A1(:,2)) A1(:,3)];B = [A(1, :) ; B1];file: /homes/leventon/Matlab/rotateMovie.m11 = linspace(0, 360);for i = 1:100view(11(i), 30);pause(0.1);endfile: /homes/leventon/Matlab/saveModelPart.mfunction saveModelpart(part, fid)r = [part{2} part{3} part{4}];fprintf(fid, `0.backslash.t%0.3f.backslash.t%0.3f.backslash.t%0.3f.backslash.n`, r);/ - subparts = part{11};n = size(subparts, 2);for i = 1:nsaveModelPart(subparts{i}, fid);endfile: /homes/leventon/Matlab/saveang.mfunction saveang(model, data, fname, deep)if (nargin < 4)deep = 1000000;endT = size(data, 2);fid = fopen(fname,`wb`);if fid==-1disp([`ERROR: fopen failed on filename: ` fname])return;end;fprintf(fid, `%d 17.backslash.n`, T);for t = i:Tmodel = setModelData(model, data(: ,t));parts = model{8};pos = zeros(1,6);for i = 2:7if (isempty(model{i}) == 0)pos(i-1) = model{i};endendn = size(parts, 2);for i = 1:nsavepart(parts{i}, pos(1:3), pos(4:6), fid, deep);endendfclose(fid);file: /homes/leventon/Matlab/savemodel.mfunction savemodel(model, fname)fid = fopen(fname,`wb`);if fid==-1disp([`ERROR: fopen failed on filename: ` fname])return;end;fprintf(fid, `17.backslash.n`);parts = model{8};n = size(parts, 2);for i = 1:nsaveModelPart(parts{i}, fid);endfile: /homes/leventon/Matlab/savepart.mfunction savepart(part, pg, og, fid, deep)if (nargin < 4)deep = 1000000;endposl = zeros(1:6);for i = 5:10if (isempty(part{i}) == 0)posl(i-4) = part{i};endendpl = posl(1:3);ol = posl(4:6);Rg = aToR(og);Rl = aToR(ol);Tg = [[Rg; 0 0 0] [diag(diag(pg)); 1]];Tl = [[R1; 0 0 0] [diag(diag(pl)); 1]];T = Tg * T1;o = rToA(T(1:3,1:3));p = T(1:3,4)';d = [T(1:3,4)' T(1:3,3)'];rx = part{2};ry = part{3};rz = part{4};T2 = T * [1 0 0 0; 0 ry/rx 0 0; 0 0 1 0; 0 0 0 1];d = [T2(1:3,1) T2(1:3,2) T2(1:3,3) T2(1:3,4)];fprintf(fid, `%0.5f %0.5f %0.5f %0.5f %0.5f %0.5f %0.5f %0.5f %0.5f$$0.2f %0.2f %0.2f.backslash.n`, d);deep = deep - 1;if (deep > 0)subparts = part{11};[lose, n] = size(subparts);for i = 1:nsavepart(subparts{i}, p, o, fid, deep);endendfile: /homes/leventon/Matlab/scl.mFOO = 1M3n4 = mencode(dance01, dance02, dance03, destroy, walkloop, runloop,fwalk, frun, thief, dance01r45, dance02r45, dance03r45, destroyr45,walkloopr45, runloopr45, fwalkr45, frunr45, thiefr4S, dance01r90,dance02r90, dance03r90, destroyr90, walkloopr90, runloopr90, fwalkr90,frunr90, thiefr90, [10 5]);FOO = 2% [u3n4, s3n4, v3n4, E3n4] = decomp(M3n4, nmk);FOO = 3% M3n4p = project(walkloop, u3n4, 100);FOO = 4M2n4 = mencode2(dance01p00, dance02p00, dance03p00, destroyp00,walkloopp00, runloopp00, fwalkp00, frunp00, thiefp00, dance01p45,dance02p45, dance03p45, destroyp45, walkloopp45, runloopp45, fwalk45,frunp45, thiefp45, dance01p90, dance02p90, dance03p90, destroyp90,walkloopp90, runloopp90, fwalkp90, frunp90, thiefp90, [10 5]);FOO = 5[u2n4, s2n4, v2n4, E2n4] = decomp2(M2n4, nmk);FOO = 6% M2n4p = project2(walkloopp45, u2n4, 50);FOO = 7[M23n4p, M2n4p] = project23(walkloopp45, M3n4, u2n4, s2n4, v2n4, 50);FOO = 8file: /homes/leventon/Matlab/sc2.mN3w = mencode(walkloop, [10 5]);M2w = mencode2(walkloopp90, [10 5]);[u2w, s2w, v2w, E2w] = decomp2(M2w, nmk);A2wp = project2(walkloopp90, u2w, 15);A23wp = project23(walkloopp90, M3w, u2w, s2w, v2w, 15);file: /homes/leventon/Matlab/sc3.m[Xn1, Mn1, un1, sn1, vn1, En1] = allbut(IS, 1, 50);[Xn2, Mn2, un2, sn2, vn2, En2] = allbut(IS, 2, 50);[Xn3, Mn3, un3, sn3, vn3, En3] = allbut(IS, 3, 50);[Xn4, Mn4, un4, sn4, vn4, En4] = allbut(IS, 4, 50);[Xn5, MnS, un5, sn5, vn5, En5] = allbut(IS, 5, 50);[Xn6, Mn6, un6, sn6, vn6, En6] = allbut(IS, 6, 50);[Xn7, Mn7, un7, sn7, vn7, En7] = allbut(IS, 7, 50);[Xn8, Mn8, un8, sn8, vn8, En8] = allbut(IS, 8, 50);[Xn9, Mn9, un9, sn9, vn9, En9] = allbut(IS, 9, 50);[Xn10, Mn10, un10, sn10, vn10, En10] = allbut(IS, 10, 50);viewaoa(Xn1, dancecl, [ ], [ ], [ ], [ ], ``, `allbut1-%03d.pgm`)viewaoa(Xn2, dancecl, [ ], [ ], [ ], [ ], ``, `allbut2-%03d.pgm`)viewaoa(Xn3, dancecl, [ ], [ ], [ ], [ ], ``, `allbut3-%03d.pgm`)viewaoa(Xn4, dancecl, [ ], [ ], [ ], [ ], ``, `allbut4-%03d.pgm`)viewaoa(XnS, dancecl, [ ], [ ], [ ], [ ], ``, `allbut5-%03d.pgm`)viewaoa(Xn6, dancecl, [ ], [ ], [ ], [ ], ``, `allbut6-%03d.pgm`)viewaoa(Xn7, dancecl, [ ], [ ], [ ], [ ], ``, `allbut7-%03d.pgm`)viewaoa(Xn8, dancecl, [ ], [ ], [ ], [ ], ``, `allbut8-%03d.pgm`))viewaoa(Xn9, dancecl, [ ], [ ], [ ], [ ], ``, `allbut9-%03d.pgm`)viewaoa(Xn10, dancecl, [ ],[ ], [ ], [ ], ``, `allbut10-%03d.pgm`)file: /homes/leventon/Matlab/sc3d.msel = seqSelect(1:10, pi/2);diff3d = seqPairDiff(sel, `3d`);save diff3d.mat diff3dFOO = 1diff2d = seqPairDiff(sel, `2d`);save diff2d.mat diff2dFDD = 2err3d = seqLeave1Out(sel, `3d`);save err3d.mat err3dFOO = 3err2d = seqLeave1Out(sel, `2d`);save err2d.mat err2dFOO = 4file: /homes/leventon/Matlab/sc4.mFOO = 1[A23n1p, A2n1p, M2n1, M3n1, u2n1, s2n1, v2n1] = allbut23(IS2, IS3, 1,2, 50);FOO = FOO + 1[A23n2p, A2n2p, M2n2, M3n2, u2n2, s2n2, v2n2] = allbut23(IS2, IS3, 2,2, SD);FOO = FDD + 1[A23n3p, A2n3p, M2n3, M3n3, u2n3, s2n3, v2n3] = allbut23(IS2, IS3, 3,2, 5D);FOO = FDD + 1[A23n4p, A2n4p, M2n4, M3n4, u2n4, s2n4, v2n4] = allbut23(IS2, IS3, 4,2, 5D);FOO = FOO + 1[A23n5p, A2n5p, M2n5, M3n5, u2n5, s2n5, v2n5] = allbut23(IS2, IS3, 5,2, SD);FOO = FOO + 1[A23n6p, A2n6p, M2n6, M3n6, u2n6, s2n6, v2n6] = allbut23(IS2, IS3, 6,2, SD);FOO = FOO + 1[A23n7p, A2n7p, M2n7, M3n7, u2n7, s2n7, v2n7] = allbut23(IS2, IS3, 7,2, 50);FOO = FOO + 1[A23n8p, A2n8p, M2n8, M3n8, u2n8, s2n8, v2n8] = allbut23(IS2, IS3, 8,2, 50);FOO = FOO + 1[A23n9p, A2n9p, M2n9, M3n9, u2n9, s2n9, v2n9] = allbut23(IS2, 1S3, 9,2, 50);FOO = FOO + 1[A23n10p, A2n10p, M2n10, M3n10, u2n10, s2n10, v2n10] = allbut23(IS2,IS3, 10, 2, 50);FOO = FOO + 1if (0)viewaoa({A23n1p, IS3{1,2}}, dancecl, [ ], [ ], [ ], [ ], ``,`allx1-%03d.ppm`)viewaoa({A23n2p, IS3{2,2}}, dancecl, [ ], [ ], [ ], [ ], ``,`allx2-%03d.ppm`)viewaoa({A23n3p, IS3{3,2}}, dancecl, [ ], [ ], [ ], [ ], ``,`allx3-%03d.ppm`)viewaoa({A23n4p, IS3{4,2}}, dancecl, [ ], [ ], [ ], [ ], ``,`allx4-%03d.ppm`)viewaoa({A23n5p, IS3{5,2}}, dancecl, [ ], [ ], [ ], [ ], ``,`allx5-%03d.ppm`)viewaoa({A23n6p, IS3{6,2}}, dancecl, [ ], [ ], [ ], [ ], ``,`allx6-%03d.ppm`)viewaoa({A23n7p, IS3{7,2}}, dancecl, [ ], [ ], [ ], [ ], ``,`allx7-%03d.ppm`)viewaoa({A23n8p, IS3{8,2}}, dancecl, [ ], [ ], [ ], [ ], ``,`allx8-%03d.ppm`)viewaoa({A23n9p, IS3{9,2}}, dancecl, [ ], [ ], [ ], [ ], ``,`allx9-%03d.ppm`)viewaoa({A23n10p, IS3{10,2}}, dancecl, [ ],[ ], [ ], [ ], ``,`allx10-%3d.ppm`)endfile: /homes/leventon/Matlab/scCore.m%%% Try to "Core" the coeffs . . .if (0)%%% Step 1 Compute the initial coeffsseqInitsel = seqSelect([1:5 7:10], pi/2);knl = seqTrain(sel, `3d`);Y = seqEncode(seqGet(6, `3d`, pi/2), {10, 5});end% nEv = 10;x0 = betai(:,1),y = Y(:,1);M = knl.enc;U = knl.U;S = sparse(knl.S);S(nEv+1:size(S,1), nEv+1:size(S,2)) = 0;S = sparse(S);V = knl.V;MtM = V*S'*S*V';Mty = V*S(1:size(S,2),1:size(S,2))'*U'*y;Im = speye(size(MtM));s2 = diag(S.*S);clear xs;pct = 1;lct = 1;for lambda = [1e3 1e4 1e5 1e6]SpLi = sparse(size.(S,2), size(S,2));SpLi(1:length(s2), 1:length(s2)) = diag(1./(s2 + lambda));SpLi = sparse(SpLi);x = x0;ct = 1;for ct = 1:5%%% Step 2. Core the coeffsex = 2;pt = max(x)*3/4;ax = abs(x);xc = ((ax >= pt) .* x) + (sign(x) .* (ax < pt) .* (ax). ex/pt (ex-1));%%% Step 3. Minimize%% E = (data - [training set] * coeffs). 2 + lambda * (coeffs -%% cored.sub.-- coeffs). 2x = (V*SpLi*V') * (Mty + lambda*xc);xs{lct, ct} = x;subplot(4,5,pct);hold off; plot(x0,`c`); hold on; plot(xc, `r`); plot(x, `b`)axis([0 300 -0.1 0.25]);axis off;plot([0 0 300 300 0], [-0.1 .25 .25 -0.1 -0.1], `k`)title(sprintf(`Lmbda=10 %d, i=%d, lct+2, ct));pct = pct + 1;pause(0.1);endlct = lct + 1;endfile: /homes/leventon/Matlab/scCorrespTest.mnumPtsAll = [1 2 5 10 20 50 75 100 200 300 400 500 600 700 740];noiseSigAll = [0 1 2 4 8 16];% numPtsAll = [1 200 740];% noiseSigAll = [4];if (0)enc2d = seqEncode(seq2d);seqMo.encoded = seqEncode(seqMo);seqFo.encoded = seqEncode(seqFo);knlMoFo = seqTrain(sel, {knlMo, knlFo}, [1 1]);seq = seqRecon(seqMo, knlMo, `3d`, 100);ErrorMo = seqError(seq, seq3d, `2d`);seq = seqRecon(seqFo, knlFo, `3d`, 100);ErrorFo = seqError(seq, seq3d, `2d`);seq = seqRecon(seqMo, seqFo, knlMoFo, `3d`, 100);ErrorMoFo = seqError(seq, seq3d, `2d`);enddatan = size(enc2d, 2);for i.sub.-- numPts = 1 :length(numPtsAll)numPts = numPtsAll(i.sub.-- numpts);zv = [ones(numpts,1); zeros(740-numpts,1)];zv = zv(randperm(740));for i.sub.-- noiseSig = 1:length(noiseSigAll)noiseSig = noiseSigAll(i.sub.-- noiseSig);STATUS = [i.sub.-- numPts i.sub.-- noiseSig]seq2dN = seqNoise(seq2d, noiseSig);enc2dN = seqEncode(seq2dN);enc2dNZ = enc2dN .* (zv * ones(1, datan));knl2dZ = knl2d;trainn = size(knl2dZ.enc{1}, 2);knl2dZ.enc{1} = knl2dZ.enc{1}.* (zv * ones(1, trainn));MM = knl2dZ.enc{1}` * knl2dZ.enc{1};[knl2dZ.U, knl2dZ.S, knl2dZ.V] = svd(MM, 0);knlMo2dZ = seqTrain(sel, {knlMo, knl2dZ}, [1 1]);knlFo2dZ = seqTrain(sel, {knlFo, knl2dZ}, [1 1]);knlMoFo2dZ = seqTrain(sel, {knlMo, knlFo, knl2dZ}, [1 1 1]);seq2dNZ = seq2d;seq2dNZ.encoded = enc2dNZ;seq1 = seqRecon(seqMo, seq2dNz, knlMo2dZ, `3d`, 100);Error{i.sub.-- numPts, i.sub.-- noiseSig, 1} = seqError(seq1, seq3d,`2d`);seq2 = seqRecon(seqFo, seq2dNz, knlFo2dZ, `3d', 100);Error{i.sub.-- numPts, i.sub.-- noiseSig, 2} = seqError(seq2, seq3d,`2d`);seq3 = seqRecon(seqMo, seqFo, seq2dNZ, knlMoFo2dZ, `3d`, 100);Error{i.sub.-- numPts, i.sub.-- noiseSig, 3} = seqError(seq3, seq3d,`2d`);seq4 = seqRecon(seq2dNZ, knl2dZ, `3d`, 100);Error{i.sub.-- numPts, i.sub.-- noiseSig, 4} = seqError(seq4, seq3d,`2d`);save errordat.mat Errorendendfile: /homes/leventon/Matlab/scFullSeqEnMin.mseq = seqGet(6, `motion`, pi/2);enc = seqEncode(seq, knlMo.encParams);nEvlnit = 10;[rcn, M3d, cc, b0, sic] = seqRecon(seq, knlMo, `3d`, nEvInit, {10 53}, knl3d.enc);options(1) = 1; options(2) = 1; options(3) = 1; options(14) = 25;global fEnergy.sub.-- Paramssnips = 1:3;clear bsfor snipnum = 0:13bs{snipnum+1,1,1} = b0(:,snips + snipnum);bs{snipnum+1,1,2} = b0(:,snips + snipnum);count = 2;b = b0(:,snips + snipnum);for nEv = [nEvInit 30 50 75 100]Si = spdiags(1./diag(knlMo.S), 0, size(knlMo.V,1), nEv);S = spdiags(doag(knlMo.S), O, nEv, size(knlMo,V,1));cc = (S * (knlMo.V'*b);ccc = reshape(cc(:,snips), lentgh(snips)*nEv, 1);M3d = knl3.enc * knlMo.V * Si;MMo = knlMo.U(:,1:nEv);fEnergyInit(MMo, M3d, enc(:,snips+snipnum), nEv, .2, 0);[e, oe, ke, lene] = fEnergy(ccc);fEnergy.sub.-- Params.1KE = oe/ke;cccm = fminu(`fEnergy`, ccc, options, `fEnergy`);b = knlMo.V * (Si * reshape(cccm, nEv, length(snips)));bs{snipnum+1, count,1} = b;bs{snipnum+1,count,2} = coreCoeff(enc(:,snips+snipnum), knlMo, b,nEv, 1000, 1);count = count + 1;save bs2.mat bsendendknlFo = seqTrain(sel, `form`);seqf = seqConvert(seqCrop(seqGet(6, `image`, pi/2), 23), `form`);[rcnf, M3df, ccf, bf, sicf] = sewrecon(seqf, knlFo, `3d`, 100, {10 53}, knl3d.enc)save form.mat ccf bf sicf% seqView(seqDecode(knl3d.enc *b, `3d`, {10, 5, 37}));file: /homes/leventon/Matlab/scManySnip.mnEv '2 100;core = {10 5 3};obsMo1 = seqCrop(seqGet(6, `motion`, pi/2), 1, 13);obsMo2 = seqCrop(seqGet(6, `motion`, pi/2),5, 18);obsMo3 = seqCrop(seqGet(6, `motion`, pi/2), 10, 23);[rec1, m, c1, b1, sic1] = seqRecon*obsMo1, knlMo, `3d`, nEv, core,knl3d.enc)[rec2, M, c2, b2, sic2] = seqRecon(obcMo2, knlMo, `3d`, nEv, core,knl3d.enc)[rec3, M, c3, b3, sic3] = seqRecon(obsMo3, knlMo, `3d`, nEv, core,kl3d.enc)file: /homes/leventon/Matlab/scMo.msel = seqselect(1:10, pi/2);diffMo6 = seqPairDiff(sel, `motion`, 6);save diffMo6.mat diffMo6FOO = 1diffMo2 = seqPairDiff(sel, `motion`, 2);save diffMo2.mat diffMo2FOO = 2errMo6 = seqLeave1Out(sel, `motion`, 6);save errMo6.mat errMo6FOO = 3errMo2 = seqLeave1Out(sel, `motion`, 2);save errMo2.mat errMo2FOO = 4diffMo3 = seqPairDiff(sel, `motion`, 3);save diffMo3.mat diffMo3FOO = 5errMo3 = seqLeave1Out(sel, `motion`, 3);save errMo3.mat errMo3FOO = 6file: /homes/leventon/Matlab/scoTest.mif (0)nEv = 50;dnsamp = 2;smth = 1;core{1} = [1e4];core{2} = 3;endif (0)obs3d = seqCrop(seqGet(6, `3d`, pi/2), 14)obsMo = seqcrop(seqGet(6, `motion`, pi/2), 14)sel = seqSelect([1:5 7:10], pi/2)knl3d = seqTrain(sel, `3d`)knlMo = seqTrain(sel, `motion`, 10, 5, 37, dnsamp, smth)end%%% (1) Rendered Original SequenceREN{1} = obsMo%%% (2) Rendered 3d recon 3d trainrec3d3d = seqRecon(obs3d, knl3d, `3d`, nEv, core, knl3d.enc)ren3d3d = seqRender(rec3d3d)REN{2} seqConvert(ren3d3d, `motion`)%%% (3) Motion recon 3d trainrecMo3d = seqRecon(obs3d, knl3d, `motion`, nEv, core, knlMo.enc);REN{3} = recMo3d%%% (4) Motion Recon Motion TrainrecMoMo = seqRecon(obsMo, knlMo, `motion`, nEv, core, knlMo.enc);REN{4} = recMoMo%%% (5) Rendered 3d Recon Motion Trainrec3dMo = seqRecon(obsMo, knlMo, `3d`, nEv, core, knl3d.enc)ren3dMo = seqRender(rec3dMo);REN{5} = seqConvert(ren3dMo, `motion`)file: /homes/leventon/Matlab/sCMultScaleEnMin.mseq = seqCrop(seqGet(6, `motion`, pi/2), 23);enc = seqEncode(seq, kn1Mo.encParams);nEvInit = 10;snips = 1:3;[rcn, M3d, cc, b, sic] = seqRecon(seq, knlMo, `3d`, nEvInit, {10 5 3},knl3d.enc);options(1) = 1; options(2) = 1; options(3) = 1; options(14) = 25;global fEnergy.sub.-- Paramsclear bsbs{1,1} = b;bs{1,2} = b;count = 2;for nEv = [nEvInit 30 50 75 100 125 150 175 200 225 250 275 294]Si = spdiags(1./diag(knlMo.S), O, size(knlMo.V,1), nEv);S = spdiags(diag(knlMo.S), O, nEv, size(knIMo.V,1));cc = (S * (knlMo.V'*b));ccc = reshape(cc(:,snips), length(snips)*nEv, 1);M3d = knl3d.enc * knlMo.V * Si;MMo = knlMo.U(:,1:nEv);fEnergyInit(MMo, M3d, enc(:,snips), nEv, .2, 0);[e, oe, ke, lene] = fEnergy(ccc);fEnergy.sub.-- Params.1KE = oe/ke;cccm = fminu(`fEnergy`, ccc, options, `fdEnergy`);b = knlMo.V * (Si * reshape(cccm, nEv, length(snips)));bs{count,1} = b;bs{count,2} = coreCoeff(enc, knlMo, b, nEv, 1000, 1);count = count + 1;save bs.mat bsendfor i = 1:10for j = [0 pi/2]seq = seqGet(1, `3d`, j);seqi = seqRender(seq);seqm = seqconvert(seqi, `motion`);for k=1:seqi.nfrainesseqi.data{k} = uint8(seqi.data{k}/4);endsave(sprintf(`Smed/seq-i-%02d-%03d.mat`, i, round(j*180/pi)),`seqm`);for k=1:seqm.nframesseqm.data{k} = uints((seqm.data{k}+pi)/(2*pi)*255);endsave(sprintf(`Smed/seq-m-%02d-%03d.mat`, i, round(j*180/pi)),`seqm`);endend% seqview(seqDecode(knl3d.enc * b, `3d`, {10, 5, 37}));file: /homes/leventon/Matlab/scNoiseScale.mif (0)% Encode all data at zero degreesM2all0 = mencode2(IS2{:,i}, [10 5]);M3all0 = mencode(IS3{:,i}, [10 5]);[u2all0, s2all0, v2all0, E2all0] = decomp2(M2all0, 37);endM3 = M3all0; u2 = u2all0; s2 = s2all0; v2 = v2all0;nEv = 50;scales = linspace(0, 20, 41);NoiseScErr = zeros(41, 10, 2);for j = 1:10for i = 1:41A2 = noisescaoa(proj2d(IS3{j}), scales(i));/ - [A23p, A2p] = project23(A2, M3, u2, s2, v2, nEv);E2 = aoaerr2(A2p, proj2d(IS3{j}));NoiseScErr(i, j, 1) = E2;E23 = aoaerr23(A23p, IS3{j});NoiseScErr(i, j, 2) = E23;STATUS = [i j E2 E23]endendfile: /homes/leventon/Matlab/scRBF.msigmas = [10 30 100 300 1000 3000];nsig = length(sigmas);/ - ncols = 4;nEv = 100;coreLambdas = [10e3 10e4];coreNIter = 3;lamba1 = 10e4;sigct '.degree.0;for sigma = sigmasrl,3 sigct = sigct + 1;subplot(1,1,1);pause(0.1);subplot(2,2,1);[alpha, D] = train.sub.-- rbf(knl3d.enc., knlFo.enc, signa, sq.sub.--dist.sub.-- fo);[y, v] = ebf(encFo(:,1), knlFo.enc, alpha, sigma);corig = (pinv(D) * v);plot(corig);title(sprintf(`RBF (signma = %d)`, sigma));axis([0 300 -0.5 0.5]);pause(0.1);print(`-dpsc`, sprintf(`Generate/rbf-fo-orig-%d.ps`, sigct));core = coreCoeff(encFo(:,1), knlFo, corig, nEv, coreLambdas,coreNIter);plot(core);title(sprintf(`Cored`, sigmas));axis([0 300 -0.5 0.5]);pause(0.1);print(`-dpsc`, sprintf(`Generate/rbf-fo-core-%d.ps`, sigct));clip = clipCoeff(encFo(:,i), knlFo, corig, nEv, coreLambdas,coreNIter);plot(clip);title(sprintf(`Clipped`, sigma));axis([0 300 -0.5 0.5]);pause(0.1);print(`-dpsc`, sprintf(`Generate/rbf-fo-clip-%d.ps`, sigct));clip1 = clip1Coeff(encFo(:,1), knlFo, corig, nEv, coreLambdas,coreNIter, lambda1);plot(clip1);title(sprintf(`Clipped, Sum=1`, sigma));axis([0 300 -0.5 0.5]);pause(0.1);print(`-dpsc`, sprintf(`Generate/rbf-fo-clipi-%d.ps`, sigct));subplot(1,1,1);s = sprintf(`Generate/rbf-fo-orig-%d-%%02d.ps`, sigct);seqview(seqDecode(knl3d.enc * corig, `3d`, {10, 5, 37}), [ ],cons(`filespec`, s));s = sprintf(`Generate/rbf-fo-core-%d-%%02d.ps`, sigct);seqview(seqDecode(knl3d.enc * core, `3d`, {10, 5, 37}), [ ],cons(`filespec`, s));s = sprintf(`Generate/rbf-fo-clip-%d-%%02d.ps`, sigct);seqview(seqDecode(knl3d.enc * clip, `3d`, {10, 5, 37}), [ ],cons(`filespec`, s));s = sprintf(`Generate/rbf-fo-clip1-%d-%%02d.ps`, sigct);seqview(seqDecode(knl3d.enc * clip1, `3d`, {10, 5, 37}), [ ],cons(`filespec`, s));endsigct = 0;for sigma = sigmassigct = sigct + 1;subplot(1,1,1);pause(0.1);subplot(2,2,1);[alpha, D] = train.sub.-- rbf(knl3d.enc, knlMo.enc, sigma, sq.sub.--dist.sub.-- mo);[y, v] = rbf(encMo(:,1), knlMo.enc, alpha, sigma);corig = (pinv(D) * v);plot(corig);title(sprintf(`HBF (sigma = %d)`, sigma));axis([0 300 -0.5 0.5]);pause(0.1);print(`-dpsc`, sprintf(`Generate/rbf-mo-orig-%d.ps`, sigct));core = coreCoeff(encMo(:,1), knlMo, corig, nEv, coreLambdas,coreNIter);plot (core);title(sprintf(`Cored`, sigma));axis([0 300 -0.5 0.5]);pause(0.1);print(`-dpsc`, sprintf(`Generate/rbf-mo-core-%d.ps`, sigct));clip = clipCoeff(encMo(:,1), knlMo, corig, nEv, coreLambdas,coreNIter);plot(clip);title(sprintf(`Clipped`, sigrna));axis([0 300 -0.5 0.5]);pause(0.1);print(`-dpsc`, sprintf(`Generate/rbf-mo-clip-%d.ps`, sigct));clip1 = clip1Coeff(encMo(:,1), knlMo, corig, nEv, coreLambdas,coreNIter, lambda1);plot(clip1);title(sprintf(`Clipped, Sum=1`, sigma));axis([0 300 -0.5 0.5]);pause(0.1);print(`-dpsc`, sprintf(`Generate/rbf-mo-clip1-%d.ps`, sigct));subplot(1,1,1);s = sprintf(`Generate/rbf-mo-orig-%d-%%02d.ps`, sigct);seqView(seqDecode(knl3d.enc * corig, `3d`, {10, 5, 37}), [ ],cons(`filespec`, s));s = sprintf(`Generate/rbf-mo-core-%d-%%02d.ps`, sigct);seqView(seqDecode(knl3d.enc * core, `3d`, {10, 5, 37}), [ ],cons(`filespec`, s));s = sprintf(`Generate/rbf-mo-clip-%d-%%02d.ps`, sigct);seqView(seqDecode(knl3d.enc * clip, `3d`, {10, 5, 37}), [ ],cons (`filespec`, s));s = sprintf(`Generate/rbf-mo-clip1-%d-%%02d.ps`, sigct);seqView(seqDecode(kn13d.enc * clip1, `3d`, {10, 5, 37}), [ ],cons(`filespec`, s));endfile: /homes/leventon/Matlab/scRenMin.mseqInitglobal fRenModelE.sub.-- Dataload Data/knlFo-090-2.matseqFo = seqGet(6, `form`, pi/2);seq3d = seqCrop(seqGet(6, `3d`, pi/2), 12);seqFoc = seqCrop(seqFo, 12);sel = seqSelect([1:5 7:10], pi/2);knl3dh = seqTrain(sel, `3d`, `headalign`, 1);[rcn3d, b3d] = seqRecon(seq3d, knl3dh, `3d`, 100, `reg`, 1e4, 1, 0, 1,1e8);[rcnFo, bFo] = seqRecon(seqFoc, knlFo, `3d`, 100, `reg`, 1e4, 1, 0, 1,1e8);c3d = knlFo.S * knlFo.V' * b3d(:,1);cFo = knlFo.S * knlFo.V' * bFo;henc = knl3dh.enc{1};nEvs = [2 5 10 15 20 25];smth = 1;dnsamp = 2;options(1) = 1; options(2) = 0.01; options(3) = 0.01; options(14) =50;cFo = 0 * c3d;cFo(1) = c3d(1);for i.sub.-- nEv=1:length(nEvs)nEv = nEvs(i.sub.-- nEv);fRenModelEcmEigInit(henc, knlFo, nEv, seqFoc, dnsamp, smth);[lose, v] = fRenModelEcmEig(c3d(1:nEv));fRenModelE.sub.-- Data.formv = v(:);cFou = fminu(`fRenModelEcmEig`, cFo(1 :nEv), options);[x3d, v3d] = fRenModelEcmEig(c3d(1:nEv));[xFo, vFo] = fRenModelEcmEig(cFo(1:nEv));[xFou, vFou] = fRenModelEcmEig(cFou);[x3d xFo xFou]Si = spdiags(1./diag(knlFo.S), 0, size(knlFo.V,1), nEv);rcn3d = seqDecode(henc * knlFo.V * Si * c3d(1:nEv), `3d`,knlFo.encParams);rcnFo = seqDecode(henc * knlFo.V * Si * cFou, `3d`,knlFo.encParams);for i = 1:10subplot(2,2,1);rgbimagesc(v3d, i);subplot(2,2,2);rgbimagesc(vFou, i);subplot(2,2,3);seqview(rcn3d, i);subplot(2,2,4);seqview(rcnFo, i);pause(0.1);print(`-dpsc`, sprintf(`minres-2-1-%02d-%d.ps`, nEv, i-1));endbFo = kniFo.V * Si * cFou;cFo = kniFo.S * kniFo.V' * bFo;endsubplot(3,1,1);rgbimagesc(v3d, 1);subplot(3,1,2);rgbimagesc(vFo);subplot(3,1,3);rgbimagesc (vFou);file: /homes/leventon/Matiab/scSpatialScale.mif (0)% Encode all data at zero degreesM2all0 = mencode2(IS2{:,1}, [10 5]);M3all0 = mencode(IS3{:,1}, [10 5]);[u2all0, s2all0, v2all0, E2all0] = decomp2(M2all0, 37);endM3 = M3all0; u2 = u2all0; s2 = s2all0; v2 = v2all0;nEv = 50;scales = logspace(-1, 1, 21);ViewErr = zeros(21, 10, 2);for j = 1:10for i = 1:21A3s = spscaoa(IS3{j}, scales(i));A2 = proj2d(A3s);[A23p, A2p] = project23(A2, M3, u2, s2, v2, nEv);E2 = aoaerr2(A2p, A2);SpScErr(i, j, 1) = E2;E23 = aoaerr23(A23p, A3s);SpScErr(i, j, 2) = E23;STATUS = [i j E2 E23]endendfile: /homes/leventon/Matlab/scTimeScale.mif (0)% Encode all data at zero degreesM2all0 = mencode2(IS2{:,1}, [10 5]);M3all0 = mencode(IS3{:,1}, [10 5]);[u2all0, s2all0, v2all0, E2all0] = decomp2(M2all0, 37);endM3 = M3all0; u2 = u2all0; s2 = s2all0; v2 = v2all0;nEv = 50;scales = logspace(-1, 1, 21);TimeScErr = zeros(21, 10, 2);for j = 1:10for i = 1:21A3s = timescaoa(IS3{j}, scales(i));A2 = proj2d(A3s);[A23p, A2p] = project23(A2, M3, u2, s2, v2, nEv);E2 = aoaerr2(A2p, A2);TimeScErr(i, j, 1) = E2;E23 = aoaerr23(A23p, A3s);TimeScErr(i, j, 2) = E23;STATUS = [i j E2 E23]endendfile: /homes/leventon/Matlab/scViewAngle.mif (0)% Encode all data at zero degreesM2all0 = mencode2(IS2{:,1}, [10 5]);M3all0 = mencode(IS3{:,1}, [10 5]);[u2all0, s2all0 v2all0] = decomp2(M2all0, 37);FOO = 1M2all0s2 = mencode2(IS2{:,[1 5]}, [10 5]);M3all0s2 = mencode(IS3{:,[1 5]}, [10 5]);[u2all0s2, s2all0s2, v2all0s2] = decomp2(M2all0s2, 37);FOO = 2M2all0s4 = mencode2(IS2{:,[1 3 5 7]}, [10 5]);M3all0s4 = mencode(IS3{:,[1 3 5 7]}, [10 5]);[u2all0s4, s2all0s4, v2all0s4] = decomp2(M2all0s4, 37);FOO = 3M2all0s8 = mencode2(IS2{:,1:8}, [10 5]);M3all0s8 = mencode(IS3{:,1:8}, [10 5]);[u2all0s8, s2all0s8, v2all0s8] = decomp2(M2all0s8, 37);FOO = 4M2all0n90 = mencode2(IS2{:,[1 3}, [10 5]);M3all0n90 = mencode(IS3{:,[1 3}, [10 5]);[u2all0n90, s2all0n90, v2all0n90] = decomp2(M2all0n90, 37);s2all0d = diag(s2all0);s2all0s2d = diag(s2all0s2);s2all0s4d = diag(s2all0s4);s2all0s8d = diag(s2all0s8);s2all0n90d = diag(s2all0n90);endnEv = 50;%% ViewErr(angle, sequence, 2dor3d, numslices)ViewErr = zeros(72, 10, 2, 5);for k = 1:5if (k == 1)M3 = M3all0; u2 = u2all0; s2 = s2all0; v2 = v2all0;elseif (k == 2)M3 = M3all0s2; u2 = u2all0s2; s2 = s2all0s2; v2 = v2all0s2;elseif (k == 3)M3 = M3all0s4; u2 = u2all0s4; s2 = s2all0s4; v2 = v2all0s4;elseif (k == 4)M3 = M3all0s8; u2 = u2all0s8; s2 = s2all0s8; v2 = v2all0s8;elseif (k == 5)M3 = M3all0n90; u2 = u2all0n90; s2 = s2all0n90; v2 =v2all0n90;endfor j = 1:100for i = -180:5:179A3r = rotaoa(IS3{j}, 1/180*pi);A2 = proj2d(A3r);[A23p, A2p] = project23(A2, M3, u2, s2, v2, nEv);E2 = aoaerr2(A2p, A2);ViewErr((1+180)/5+1, j, 1, k) = E2;E23 = aoaerr23(A23p, A3r);ViewErr((1+180)/5+1, j, 2, k) = E23;STATUS = [i j k E2 E23]endendendfile: /homes/leventon/Matlab/segtorect3.mfunction r3 = segtorect3(p1, p2, rad)% Take two 3d points and a radius and make a rectangle% between them.c = (p1+p2)/2;z = (p1-p2)/2;file: /homes/leventon/Matlab/selectMarker.m% function which = selectMarker(M3d, MoldRep, obs, betai, offset, t)t = 10.offset = seqMo.head.dat.sub.-- d;obs = {seqMo};nEv = 100;knl = knlMo;MtM = knl.U * knl.S * knl.V';lambda = 1e3;lambdal = 1e9;lambdac = 1e6;rcn2 = seqDecode(M3d * betai, `3d`, knl.encParams, seqMo);seqView2({seqIm, rcn2}, t);d = M3d * betai;d = reshape(d, 3, 37, 10);d = d(:,:,t)';woff = camToWorld(offset);d(:,1) = d(:, 1) + woff(1,1);d(:,3) = d(:, 3) + woff(1,3);d2 = worldToCam(d);NumTrainRep = length(obs);if (0)for i = 1:NumTrainRepMobs{i} = seqEncode(obs{j}, knl.encParams);endMty = 0;for i = 1:NumTrainRepMty = Mty + knl.weights(i) * knl.enc{i}' * Mobs{i};endendp = ginput(1);CCterms = 0;CtPterms = 0;while (p(1) > 0)[v, which] = min(sum((d2-ones(37,1)*p). 2, 2));hold onplot(d2(which, 1), d2(which, 2), `y*`);pnew = ginput(1);pnewW = camToWorld(pnew)';plot([d2(which, 1) pnew(1)], Ld2(which, 2) pnew(2)], `yo-`);Co = sparse([1 3], [(which-1)*3+1+(t-1)*37*3,(which-1)*3+3+(t-1)*37*3], [1 1], 3, 1110);C = Co * henc;CtC = C'*C;pneww = (camToWorld(pnew) - woff(1,:))';CCterms = CCterms + lambda * CtC;CtPterms = CtPterms + C' * (lambda * pnewW);p = ginput(1);endMtMpLCtCi = pinv(MtM + CCterms + lambda1);x = MtMpLCtCi * (Mty + CtPterms + lambda1);ex = 2;pt = max(x)*3/4;% Computing the coring of the coefficientsax = abs(x);xc = x;xc = ((ax >= pt) .* x) + (sign(x) .* (ax < pt) **(ax). ex/pt (ex-1));xc = xc .* (xc > 0);MtMpLCtCi = pinv(MtM + CCterms + lambda1 + lambdac * eye(size(MtM)));x = MtMpLCtCi * (Mty + CtPtems + lambda1 + lambdac * xc);betai2 = x;rcn2 = seqDecode(M3d * betai2, `3d`, knl.encParams, seqMo);seqView2({seqIm, rcn2}, t);file: /homes/leventon/Matlab/seqConvert.mfunction seqn = seqConvert(seq, rep, varargin)% seqConvert Convert a Sequence from One Representation to Another%% seqNew = seqConvert(seq, rep) converts the sequence%`seq`% into the new representation `rep`%% seqNew = seqConvert(seq, rep, theta) rotates the% sequence by theta before converting.%% seqNew = seqConvert(seq, rep; {theta}, {pps}) uses pps%% pulses per segment.% Valid Conversions are as follows:%% `3s` to `3d`, `2d`, `stick`, `ang`, `pulse` (*)% `2d` to `stick`, `pulse`% `image` to `form`, `edge`, `motion`% `pulse` to `formc`, `stackedpulse`, `stackedpulse.sub.-- no.sub.-- normals`% `stackedpulse`to `pulse`, `formc`, `derivs`% `stackedpulse.sub.-- no.sub.-- normals`to `pulse`, `formc`, `derivs`%%% To convert from `3d` to `image`, `form`, `edge`, or `motion`,% first use seqRender to convert it to `image`.%% 10/9/97 wtf added conversion from 2d to pulse.% to convert from 2d to pulse, at 2 pulses per line segment,% viewed from theta = 0, for a sequence of 20 frames:% seqpulse = seqconvert(seqCrop(seq2d,20), `pulse`, {0}, {2});% Added varargin2.% To do: clean up this varargin1, vargagin2 call, with the cells.% Why are cells needed for pps?% 10/10/97 added pulse to formb conversion. formb is sensor% responses, stacked in a form convenient for me to work with% for right now, and combined with derivative information.% For now, formb assumes only one contrast sensor,% not an array of different sensors at different% orientations.% 10/15/97 wtf added `stackedpulse`. that is all the pulse info,% stacked into one tall column vector. Its purpose:% for taking linear combinations of sequences in pulse rep.% To do: omit the orientation component??% 10/19/97 wtf removed formb. added pulse to formc conversion.% (formb rep. had both sensor response and derivs. formc% just has sensor response. derivs rep. has derivatives.)global seqINFO%% this handles conversion of multiple sequences at once (cell-list of%% seqs)if (iscell(seq))n = length(seq);seqn = cell(size(seq));for i=1:nseqn{i} = seqconvert(seq, rep, varargin{1:length(varargin)});endreturnendoldtype = seq.type;newtype = rep;switch oldtypecase `3d`switch newtypecase `3d`seqn = seq;case `2d` seqn = seq;seqn.type = `2d`;A3 = seqn.data;n = size(A3, 1);A3a = A3(2:n, :);A2a = [A3a(:,1) A3a(:,3)];A2 = [A3(1, 1:2); A2a];seqn.data = A2;case `stick` seqn = seq; seqn.data = stickify(seq); seqn.type = `stick`; seqn.name = [`Stick ` seq.name];case `image` error(`Use seqRender to do this`);case `motion`error(`Use seqRender to convert to image then Matlab for image->motion`);case `form`error(`Use seqRender to convert to image then Matlab for image->form`);case `ang` if (seq.subtype == 1) seqn = seq;seqn.type = `ang`;seqn.data = mkr2ang(seq.data); seqn.model = seqINFO.repAngle.model; elseerror(`Can only convert subtype 1 to angle representation`); endcase `pulse` seqn = seqconvert(seqconvert(seq, `2d`, varargin{:}), . . .`pulse`, varargin{:});case `formc` seqn = seqConvert(seqConvert(seq, `pulse`), . . .`formc`, varargin{:});otherwiseerror(`Bad Representation`);endcase `2d`switch newtypecase `2d`seqn = seq;case `stick` seqn = seq; seqn.data = stickify(seq);case `pulse` seqn = seq; seqn.type = `pulse`; if (nargin < 3)pps = 1;elsepps = varargin{1};end seqn.data = pulseify(seq, pps);case `formc` seqn = seqConvert(seqConvert(seq, `pulse`), . . .`formc`, varargin{:});otherwiseerror(`Cannot convert from 2d to anything`);endcase `pulse`switch newtypecase `formc` seqn = seq; seqn.type = `formc`;% setup blank array for data. {frame, 1} is sensor response;seqn.data = cell(seq.nframes, 1); if (nargin < 3)error(`need sensor.sub.-- pos, sensor.sub.-- ori, twosigmasq, cos.sub.--power arguments`);returnelsesensor.sub.-- pos = varargin{1};sensor.sub.-- ori = varargin{2};twosigmasq = varargin{3};cos.sub.-- power = varargin{4};end% load data with cells (one per time frame) of sensor response info for frame = 1:seq.nframes% Subtracting the head position here. headaln = seq.data{frame,1}- . . .ones(size(seq.data{frame,1},1), 1) *seq.head.data2d(1, :); seqn.data{frame,1} = . . . getSR(sensor.sub.-- pos, sensor.sub.-- ori, headaln',seq.data{frame,2}', . . .twosigmasq, cos.sub.-- power);end seqn.sensor.sub.-- pos = sensor.sub.-- pos; seqn.sensor.sub.-- ori = sensor.sub.-- ori; seqn.twosigmasq = twosigmasq; seqn.cos.sub.-- power = cos.sub.-- power;case `stackedpulse` seqn = seq; seqn.type = `stackedpulse`;seqn.data = [ ]; for frame = 1:seq.nframes seqn.data = [seqn.data; . . . reshape(seq.data{frame,1}', size(seq.data{frame,1},1)*2,1); . . . reshape(seq.data{frame,2}), size(seq.data{frame,2},1)*2,1)]; endcase `stackedpulse no.sub.-- normals` seqn = seq; seqn.type = `stackedpulse.sub.-- no.sub.-- normals';seqn.data = [ ]; for frame = 1:seq.nframes seqn.data = [seqn.data; . . . reshape(seq.data{frame,1})', size(seq.data{frame,1},1)*2,1)]; endendcase `stackedpulse`switch newtypecase `pulse` seqn = seq; seqn.type = `pulse`; seqn.data = cell(seq.nframes,2); [col, lose] = size(seq.data); npts = col/(4*seq.nframes); tmp = zeros(npts*4,1); for frame = 1:seq.nframes tmp = seq.data((frame-1)*npts*4+(1:npts*4),1);seqn.data(frame,1) = {reshape(tmp(1:2*npts),2,npts))};seqn.data(frame,2) = {reshape(tmp((2*npts+1):4*npts),2,npts)'};end% formc is the response to spatial information. not derivatives.case `formc` seqn = seq; seqn.type = `formc`;% setup blank array for data. {frame, 1} is sensor response; seqn.data = cell(seq.nframes,1); if (nargin < 3)error(`need sensor.sub.-- pos, sensor.sub.-- ori, twosigmasq, cos.sub.--power arguments`);returnelsesensor.sub.-- pos = varargin{1};sensor.sub.-- ori = varargin{2};twosigmasq = varargin{3};cos.sub.-- power = varargin{4};end[col, lose] = size(seq.data);npts = col/(4*seq.nframes);tmp = zeros(npts*4,1);for frame = 1:seq.nframes tmp = seq.data((frame-1)*npts*4+(1:npts*4) ,1);%% Subtracting the head position here. headaln = reshape(tmp(1:2*npts),2,npts)'; headaln = headaln - . . .ones(size(headaln, 1), 1) * seq.head.data2d(1,:); seqn.data{frame,1} = . . . getSR(sensor.sub.-- pos, sensor.sub.-- ori, headaln', . . reshape(tmp((2*npts+1):4*npts),2,npts),. . .twosigmasq, cos.sub.-- power);endseqn.sensor.sub.-- pos = sensor.sub.-- pos;seqn.sensor.sub.-- ori = sensor.sub.-- ori;seqn.twosigmasq = twosigmasq;seqn.cos.sub.-- power = cos.sub.-- power;% derivs are the derivatives w.r.t. pulse parameters.case `derivs`seqn = seq;seqn.type = `derivs`;% setup blank array for data. {frame, 2} are sensor derivatives;seqn.data = cell(seq.nframes,1);if (nargin < 3)error(`need sensor.sub.-- pos, sensor.sub.-- ori, twosigmasq, cos.sub.--power arguments`);returnelsesensor.sub.-- pos = varargin{1};sensor.sub.-- ori = varargin{2};twosigmasq = varargin{3};cos.sub.-- power = varargin{4};end[col, lose] = size(seq.data);npts = col/(4*seq.nframes);tmp = zeros(npts*4,1);for frame = 1:seq.nframes tmp = seq.data((frame-1)*npts*4+(1:npts*4),1); seqn.data{frame,1} = . . .getSD(sensor.sub.-- pos, sensor.sub.-- ori, . . . reshape(tmp(1:2*npts),2,npts), . . .reshape(tmp((2*npts+1):4*npts),2,npts), . . . twosigmasq, cos.sub.-- power);endendcase `stackedpulse.sub.-- no.sub.-- normals`switch newtypecase `pulse` seqn = seq; seqn.type = `pulse`; seqn.data = cell(seq.nframes,2); [col, lose] = size(seq.data); npts = col/(2*seq.nframes); tmp = zeros(npts*2,1); for frame = 1:seq.nframes tmp = seq.data((frame-1)*npts*2+(1:npts*2),1);seqn.data(frame,1) = {reshape(tmp(1:2*npts),2,npts)'}; %% just make up some orientation data, so that sizes match.seqn.data(frame,2) = {sqrt(2)*ones(npts,2)};end% formc is the response to spatial information.not derivatives.case `formc` seqn = seq; seqn.type = `formc`;% setup blank array for data. {frame, 1} is sensor response; seqn.data = cell(seq.nframes,1); if (nargin < 3)error(`need sensor.sub.-- pos, sensor.sub.-- ori, twosigmasq, cos.sub.--power arguments`);returnelsesensor.sub.-- pos = varargin{1};sensor.sub.-- ori = varargin{2};twosigmasq = varargin{3};cos.sub.-- power = varargin{4};end [col, lose] = size(seq.data);colseq.nframes npts = col/(2*seq.nframes); tmp = zeros(npts*2,1); for frame = 1:seq.nframes tmp = seq.data((frame-1)*npts*2+(1:npts*2),1);%% Subtracting the head position here. headaln = reshape(tmp(1:2*npts),2,npts)'; headaln = headaln - . . .ones(size(headaln, 1), 1) * seq.head.data2d(1,:); seqn.data{frame,1} = . . . getSR(sensor.sub.-- pos, sensor.sub.-- ori, headaln', . . . sqrt(2)*ones(2,npts), . . . twosigmasq, cos.sub.-- power);end seqn.sensor.sub.-- pos = sensor.sub.-- pos; seqn.sensor.sub.-- ori = sensor.sub.-- ori; seqn.twosigmasq = twosigmasq; seqn.cos.sub.-- power = cos.sub.-- power;% derivs are the derivatives w.r.t. pulse parameters.case `derivs` seqn = seq; seqn.type = `derivs`;% setup blank array for data. {frame, 1} are sensor derivatives; seqn.data = cell(seq.nframes,1); if (nargin < 3)error(`need sensor.sub.-- pos, sensor.sub.-- ori, twosigmasq, cos.sub.--powerarguments`);returnelsesensor.sub.-- pos = varargin{1};sensor.sub.-- ori = varargin{2};twosigmasq = varargin{3};cos.sub.-- power = varargin{4};end [col, lose] = size(seq.data); npts = col/(2*seq.nframes); tmp = zeros(npts*2,1); for frame = 1:seq.nframes tmp = seq.data((frame-1)*npts*2+(1:npts*2),1); seqn.data{frame,1} = . . .getSD.sub.-- no.sub.-- normals(sensor.sub.-- pos, sensor.sub.-- ori, . .. reshape(tmp(1:2*npts) ,2,npts), . . . sqrt(2)*ones(2,npts), . . . twosigmasq, cos.sub.-- power);endendcase `image`switch newtypecase `image`seqn = seq;case `form` seqn = seq; nfr = length(seq.data); for t = 1:nfr datan{t} = compForm(seq.data{t}, pi/4*[0:3]); endseqn.data = datan;seqn.type = `form`;seqn.params.angles = pi/4*[0:3];%% sensor reponses with sensor positions and orientations%% explicitly described,case `formc` seqn = seq; nfr = length(seq.data); if (nargin < 3)error(`need sensor.sub.-- pos, sensor.sub.-- ori, twosigmasq, cos.sub.--power arguments`);returnelsesensor.sub.-- pos = varargin{1};sensor.sub.-- ori = varargin{2};twosigmasq = varargin{3};cos.sub.-- power = varargin{4};end% setup blank array for data. {frame, 1} is sensor response;seqn.data = cell(seq.nframes,1);sensor.sub.-- pos = sensor.sub.-- pos + . . .seq.head.data2d(1, :)' * ones(1, size(sensor.sub.-- pos,2));for t = 1:nfr seqn.data{t,1} = compFormc(seq.data{t}, 0, . . . sensor.sub.-- pos, sensor.sub.-- ori, twosigmasq, cos.sub.-- power);endseqn.type = `formc`;seqn.params.angles = 0;seqn.sensor.sub.-- pos = sensor.sub.-- pos;seqn.sensor.sub.-- ori = sensor.sub.-- ori;seqn.twosigmasq = twosigmasq;seqn.cos.sub.-- power = cos.sub.-- power;case `edge` seqn = seq; nfr = length(seq.data); for t = 1:nfr seqn.data{t} = call.sub.-- canny(upscale(seq.data{i}, 2), 2, 2, 5); endseqn.type = `edge`;case `motion` seqn = seq;nfr = length(seq.data);for t = 1:nfr-1 tdatan{t} = .compNorVel4(seq.data{t:t+1}, [0 pi/4 pi/23*pi/4], 0.1);endseqn.data = datan;seqn.type = `motion`;seqn.params.angles = [0 pi/4 pi/2 3*pi/4];seqn.params.thresh = 0.1; seqn.nframes = nfr-1;otherwiseerror(`Can only convert from image to motion`);endcase `motion`switch newtypecase `motion`seqn = seq;otherwiseerror(`Cannot convert from motion to anything`);endcase `ang`switch newtypecase `ang`seqn = seq;otherwiseerror(`Cannot convert from angle to anything');endcase `form`switch newtypecase `form`seqn = seq;otherwiseerror(`Cannot convert from form to anything');endotherwiseerror(`Bad Representation`);endfile: /homes/leventon/Matlab/seqCrop.mfunction seqc = seqCrop(seq, mn, mx)% seqCrop Crop a sequence%% seqc = seqCrop(seq, nfr) crops out the first nfr frames of the% sequence seq.%% seqc = seqCrop(seq, first, last) crops out the frames of the% sequence between #first and #last (inclusive).%% BUGS: Doesn't completely work for all representations yet.% oct. 15, 1997 wtf added pulse representation.if (nargin < 3)mx = mn;mn = 1;endglobal seqINFOnmkr = seqINF0.nurnmkr(seq.subtype);switch seq.typecase {`3d`, `3dc`, `2d'}seqc = seq;seqc.nframes = mx-mn+1;nb = (mn-1) * nmkr + 2;ne = nmkr*mx + 1;seqc.data = [seqc.data(1, :); seqc.data(nb:ne,:)];seqc.data(1,1) = mx-mn+1;seqc.name = [`Cropped ` seq.name];seqc.head = seqCrop(seq.head, mn, mx);case {`image`, `motion`, `form`, `edge`}seqc = seq;seqc.nframes = mx-mn+1;seqc.data = {seqc.data{mn:mx}};seqc.name = [`Cropped ` seq.name];seqc.head = seqcrop(seq.head, mn, mx);case {`pulse`}seqc = seq;seqc.nframes = mx-mn+1;seqc.data = seqc.data(mn:mx,:);seqc.name = [`Cropped ` seq.name];seqc.head = seqCrop(seq.head, mn, mx);case {`head`}seqc = seq;seqc.nframes = mx-mn+1;seqc.data2d = seq.data2d(mn:mx, :);if (exists(seq, `data3d`)) seqc.data3d = seq.data3d(mn:mx, :);endseqc.name = [`Cropped ` seq.name];case `ang`endfile: /homes/leventon/Matlab/seqData.mfunction d = seqData(s)seqData Accessor function returning the raw sequence data%% d = seqData(seq) returns the raw data of the sequence seq.d = s.data;file: /homes/leventon/Matlab/seqDecode.mfunction seq = seqDecode(M, rep, params, seq, meanval)% seqDecode Decode raw data into a sequence%%%global seqINFOif (nargin < 4)offsets = [ ];disp(`Warning: assuming subtype 1`);seq.subtype = 1;elseoffsets = seq.head.data2d;endif (nargin < 5), meanval = zeros(size(M,1), 1); endif ((exists(params, `submean`)==0) .vertline. (params.submean == 0))meanval = zeros(size(M,1), 1);endif (params.headalign == 0)offsets = [ ];end[nrow, ncol] = size(M);switch repcase `3d`s = params.fps;k = params.step;nmk = seqINFO.nummkr(seq.subtype);nRows = nmk * 3;nfr = (ncol-1)*k+s;MA = zeros(nmk*3*nfr, ncol);MC = zeros(nfr, 1);x = linspace(-(s-1)/2, (s-1)/2, s)`;sig = s/2-1;mask = exp(-x.*x/(sig*sig));% mask = ones(s,1);% mask = [ones(s/2,1); 0.000001*ones(s/2,1)];M = M + meanval*ones(1, size(M, 2));MM = zeros(3, nmk, s, ncol);for i = 1:ncolb = 1;for j = 1:s for kk = 1:nmk MM(:, kk, j, i) = M(b:b+2, i); b = b + 3; endendendif (isempty(offsets) == 0)%% Hack. hack. Camera Coords to Worldp = camToWorld(offsets);for i=1:ncol t = (i-1)*k+1;MM(1, :, :, i) = MM(1, :, :, i) + p(t,1);MM(3, :, :, i) = MM(3, :, :, i) + p(t,3);endendMA = zeros(nmk*3*nfr, 1);cols = [ ];for t = 1:nfrif (mod(t-1, k) == 0)newcol = (t-1)/k+1;if (newcol <= ncol) if (isempty(cols) == 0) ycom = 0;msum = 0; for i = cols j = t - (i-1)*k; ycom = ycom + mask(j) * sum(MM(2, :, j, i),2)/nmk; msum = msum + mask(j); end ycom = ycom / msum; thisycom = sum(MM(2, :, 1, newcol), 2)/nmk; MM(2, :, :, newcol) = MM(2, :, :, newcol) - thisycom +ycom; end cols = [cols newcol];endendmsum = 0;ycom = 0;cols2 = cols;for i = cols2j = t - (i-1)*k;if (j == s) n = length(cols); cols = cols(2:n);endb = (t-1)*nmk*3 + 1;MA(b:b+nRows-1) = MA(b:b+nRows-1) + mask(j) *vectorize(MM(:, :, j, i)');msum = msum + mask(j);endMA(b:b+nRows-1) = MA(b:b+nRows-1) / msum;endA = [nfr nmk 30; matricize(MA, 3)];seq.type = `3d`;seq.name = `Reconstructed`;seq.data = A;seq.angle = 0;seq.nframes = nfr;case `2d`s = params.fps;k = params.step;nmk = seqINFO.nummkr(seq.subtype);nRows = nmk * 2;nfr = (ncol-1)*k+s;MA = zeros(nmk*2*nfr, 1);MC = zeros(nfr, 1);x = linspace(-(s-1)/2, (s-1)/2, s)';sig = s/2-1;mask = exp(-x.*x/(sig*sig));% mask = ones(s,1);b = 1;for i = 1:ncolMC(b:b+s-1) = MC(b:b+s-1) + mask;for j = 1:sb2 = (b - 1) * 2*nmk + 1;e2 = b2 + nRows - 1;MA(b2:e2) = MA(b2:e2) + mask(j) * M((j-1)*nRows+1:j*nRows, i);b = b + 1;endb = b + -s + k;endfor i = 1:nfrb = (i-1)*nmk*2 + 1;e = i*nmk*2;MA(b:e) = MA(b:e) / MC(i);endA = [nfr nmk; matricize(MA, 2)];seq.type = `2d`;seq.name = `Reconstructed`;seq.data = A;seq.angle = 0;seq.nframes = nfr;case `image`case `motion`xxx = (reshape(M, 20, 20, 4, 10));for i = 1:10for j = 1:4data{i}(:,:,j) = xxx(:,:,j,i)';endendseq.type = `motion`;seq.name = `Reconstructed`;seq.data = data;seq.angle = 0;seq.nframes = 10;case `form`seq.data = [ ];case `ang`otherwiseerror(`Bad Representation`);endfile: /homes/leventon/Matlab/seqDownsample.mfunction seq = seqDownsample(seq, times)% seqDownsample Downsample an image sequenceif (nargin < 2), times = 1; endn = seq.nframes;switch seq.typecase {`image`, `motion`, `form`, `stick`}for i = 1:nfrseq.data{i} = downsample(seq.data{i}, times);endseq.head.data2d = seq.head.data2d / (2 times);otherwiseerror(`Cannot downsample this type of sequence`);endseq.name = `Downsampled ` seq.name];if (exists(seq, `scale`))seq.scale = seq.scale / (2 times);elseseq.scale = 1 / (2 times);endfile: /homes/leventon/Matlab/seqEdge.mfunction seqEd = seqEdge(seqIm, smth, cannyl, canny2, supersamp)% SeqEdge Canny Edge Detect a sequence%% seqEd = seqEdge(seqIm) edge detects the sequence%% Note: There are parameters here that I didn't have time to make as% optionsscale = 1;if (exists(seqIm, `scale`)), scale = seqIm.scale; endif (nargin < 2), smth = 2; endif (nargin < 3), cannyl = 3; endif (nargin < 4), canny2 = 6; endif (nargin < 5), supersamp = 2; endnfr = seqIm.nframes;seqEd = seqIm;for i = 1:nfrON.sub.-- FRAME = iI = upsample(seqIm.data{i}, supersamp);I = call.sub.-- canny(I, smth, canny1, canny2);seqEd.data{i} = uint8(J.sup..about. =0);endseqEd.name = `Edge Detected ` seqIm.name];seqEd.type = `edge`;seqEd.scale = supersamp * scale;file: /homes/leventon/Matlab/seqEncode.m% seqEncode Encode a Sequence as a Matrix of snippets%% seqEncode(seq, . . . options . . .%% Options:% `fps`, fps "Frames Per Snippet" Default = 10% `step`, step "Step sizes between Snippets" Default = 5% `smooth`, sm "Amount of Smoothing" Default = 0% `downsample`, dnsamp "Amount of Downsampling" Default = 2% `headalign`, headalign "Whether or not to zero the head%position"% Default = 1 when rep == "image","motion","form"% Default = 0 otherwise% `force`, force "Force encoding . . ."function M = seqEncode(seq, varargin)global seqINFOif (iscell(seq) == 0)seq = {seq};endheadalign = 0;for i = 1:length(seq)switch seq{i}.typecase {`image`, `motion`, `form`}headalign = 1;endendoptions = args(varargin{1 :length(varargin)});options = args(options, `fps`, 10, `step`, 5, `smooth`, 0,`downsample`, 2, `headalign`, headalign,. `force`, 0, `normals`, 0);s = options.fps;k = options.step;dnsamp = options.downsample;smth = options.smooth;nseq = length(seq);rep = seq{1}.type;if (seqINFO.defaults.debug)disp(sprintf(`Encoding seq %s of rep %s . . .`, seq{1}.name, rep));endswitch repcase {`3d`, `2d`}if (options.headalign == 0)if (rep == `3d`) ndim = 3;else ndim = 2;endnmk = seqINFO.nummkr(seq{1}.subtype);nRows = nmk * s * ndim;c = 1;for w = 1:nseqA = seq{w}.data;nfr = A(1,1);nmkchk = A(1,2);if (nmkchk .sup..about. = nmk) error(`Bad AOA Matrix!`);end[nrowA, ncolA] = size(A);if (ncolA .sup..about. = ndim) error(`Bad AOA Matrix!`);endA = A(2:nrowA, :);ndat = nfr * nmk;if (options.force)last = (nfr-s+1);elselast = (nfr-s-1);endfor i = 1:k:last b = (i-1)*nmk+1; e = b + nmk * s - 1; v = vectorize(A(b:e, :)); M(1:nRows,c) = v; c = c + 1;endendelseFOO = 1;if (rep == `3d`)ndim = 3;elsendim = 2;endnmk = seqINFO.nummkr(seq{1}.subtype);nRows = nmk * s * ndim;c = 1;for w = 1:nseqA = seq{w}.data;nfr = A(1,1);nmkchk = A(1,2);if (nmkchk .sup..about. = nmk) error(`Bad AOA Matrix!`);end[nrowA, ncolA] = size(A);if (ncolA .sup..about. = ndim) error(`Bad AOA Matrix!`); end A = A(2:nrowA, :); ndat = nfr * nmk;if (options.force)last = (nfr-s+1); elselast = (nfr-s-1); end for i = 1:k:last b = (i-1)*nmk+1; e = b + nmk * s - 1; pos = seq{w}.head.dat.sub.-- d(i,:); v = vectorize(A(b:e, :) - ones(nmk*s,1)*pos); M(1:nRows,c) = v; c = c + 1; endendendcase `image`M = [ ];for w = 1:nseqENCODING.sub.-- SEQ = wdata = seq{w}.data;nfr = length(data);for i=1:k:(nfr-s) thissnip = [ ]; posx = seq{w}.head.data2d(i,1); posy = seq{w}.head.data2d(i,2); for t=((1:s)+i-1) Ic = imCrop(data{t}, round(posy)-40, round(posx)-40, 80,80); Icp = downsample(Ic, dnsamp); thissnip = [thissnip; vectorize(Icp)]; end M = [M thissnip];endend`motion`M = [ ];for w = 1:nseqENCODING.sub.-- SEQ = [w dnsamp smth]data = seq{w}.data;nfr = length(data);for i=1:k:(nfr-s) thissnip = [ ]; posx = seq{w}.head.data2d(i,1);posy = seq{w}.head.data2d(i,2);for t=((1:s)+i-1) Ic = imCrop(data{t}, round(posy)-40, round(posx)-40, 80,80); Icp = downsample(Ic, dnsamp);Icp = smooth(Icp, smth); for d = 1:4 STATUS = [w i t d]; thissnip = [thissnip; vectorize(Icp(:,:,d))]; endendM = [M thissnip];endendcase `form`M = [ ]; - for w = 1:nseqENCODING.sub.-- SEQ = [w dnsamp smth]data = seq{w}.data;nfr = length(data);if (options.force)last = (nfr-s+1);elselast = (nfr-s-1);endfor i = 1:k:lastthissnip = [ ];posx = seq{w}.head.data2d(i,1);posy = seq{w}.head.data2d(i,2);for t=((1:s)+i-1) Ic = imCrop(data{t}, round(posy)-40, round(posx)-40, 80,80); Icp = downsample(Ic, dnsamp);Icp = smooth(Icp, smth); for d = 1:4 STATUS = [w i t d]; thissnip = [thissnip; vectorize(Icp(:, :,d))]; endendM = [M thissnip];endendcase `pulse`M = [ ];for w = 1:nseqdata = seq{w}.data;nfr = seq{w}.nframes;if (options.force)last = (nfr-s+1);elselast = (nfr-s-1);endfor i = 1:k:last thissnip = [ ]; for t=((1:s)+i-1) thissnip = [thissnip; seq{w}.data{t,1}(:)]; if (options.normals) thissnip = [thissnip; seq{w}.data{t,2}(:)]; end end M = [M thissnip];endendcase `stick`M = [ ];for w = 1:nseqENCODING.sub.-- SEQ = [w dnsamp smth]data = seq{w}.data;nfr = seq{w}.nframes;if (options.force)last = (nfr-s+1);elselast = (nfr-s-1);endfor i = 1:k:last thissnip = [ ]; posx = seq{w}.head.data2d(i,1); posy = seq{w}.head.data2d(i,2); for t=((1:s)+i-1) Ic = imCrop(data{t}, round(posy)-40, round(posx)-40, 80,80); Icp = downsample(Ic, dnsamp);Icp = smooth(Icp, smth); for d = 1:4 STATUS = [w i t d]; thissnip = [thissnip; vectorize(Icp(:,:d))]; end end M = [M thissnip];endendcase `ang`M = aencode(seq{i}.data, [10, 5]);otherwiseerror(`seqEncode: Bad Representation`);endif (isempty(M))disp(`seqEncode: WARNING: Returning empty encoded matrix`);disp(` Perhaps (num.sub.-- time.sub.-- steps.sub.-- per.sub.-- snippet >num.sub.-- time.sub.-- steps) ?`);endif (seqINFO.defaults.debug)disp(sprintf(`Done`));endfile: /homes/leventon/Matlab/seqError.mfunction [err, seqicen, errvals] = seqError(seq1, seq2, rep, cenalign)% seqError Compute the 2d error between to sequences%% err = seqError(seq1, seq2) returns the rms error between the two% seqs.%errvals = [ ];if (nargin < 3), rep = `2d`; endif (nargin < 4), cenalign = 0; endswitch repcase `3d`seq1a = seqConvert(seq1, `3d`);seq2a = seqconvert(seq2, `3d`);A = seq1a.data;B = seq2a.data;n = size(A, 1);nmk = A(1,2);nfr = A(1,1);A = A(2:n,:);B = B(2:n,:);if (cenalign == 0)err = mean(sqrt(sum((A-B). 2, 2)));elseif (cenalign == 1)A = permute(reshape(A, nmk, nfr, 3), [1 3 2]);B = permute(reshape(B, nmk, nfr, 3), [1 3 2]);Am = mean(A, 1);Bm = mean(B, 1);err = 0;for t = 1:nfr At = A(:,:,t) - ones(nmk,1)*Am(:,:,t); Bt = B(:,:,t) - ones(nmk,1)*Bm(:,:,t); C(:,:,t) = At + ones(nmk,1)*Bm(:,:,t); err = err + mean(sqrt(sum((At-Bt). 2, 2)));errvals = [errvals sqrt(sum((At-Bt). 2, 2))];endC = reshape(permute(C, [1 3 2]), nmk * nfr, 3);seq1cen = seq1;seq1cen.data = [seq1a.data(1,:); C];err = err / nfr;elseif (cenalign == 2)A = permute(reshape(A, nmk, nfr, 3), [1 3 2]);B = permute(reshape(B, nmk, nfr, 3), [1 3 2]);Am = mean(mean(A, 1), 3);Bm = mean(mean(B, 1), 3);err = 0;for t = 1:nfr At = A(:,:,t) - ones(nmk,1)*Am(:,:); Bt = B(:,:,t) - ones(nmk,1)*Bm(:,:); C(:,:,t) = At + ones(nmk,1)*Bm(:,:); err = err + mean(sqrt(sum((At-Bt). 2, 2)));errvals = [errvals sqrt(sum((At-Bt). 2, 2))];endC = reshape(permute(C, [1 3 2]), nmk * nfr, 3);seq1cen = seq1;seq1cen.data = [seq1a.data(1,:); C];err = err / nfr;endcase `2d`seq1a = seqConvert(seq1, `2d`);seq2a = seqConvert(seq2, `2d`);A = seq1a.data;B = seq2a.data;n = size(A, 1);A = A(2:n,:);B = B(2:n,:);err = mean(sqrt(sum((A-B). 2, 2)));otherwiseerror(`Not Implemented`);endfile: /homes/leventon/Matlab/seqExampleTest.m% seqTest Script with some test examplesseqInitseqView(seqGet(1));seqView(seqGet(1, `2d`));seqView(seqGet(1, `2d`, pi/2));seqView(seqGet(1, `image`));sel = seqSelect([1:5 7:10]);knl = seqTrain(sel, `2d`);obs = seqGet(6);rcn = seqRecon(obs, knl);seqView(rcn);sel = seqSelect([1:5 7:10]);knl = seqTrain(sel, `2d`);obs = seqGet(6, `2d`);rcn = seqRecon(obs, knl, `3d`);seqView(rcn);sel = seqSelect(6);knl = seqTrain(sel, `motion`);obs = seqGet(6, `motion`);rcn = seqRecon(obs, knl, `3d`);seqView(rcn);file: /homes/leventon/Matlab/seqExamples.mfunction seqExamples (which)% seqExamples Show some examples on how to use the code%% seqExamples(which) goes through example #which.%% (**) Be sure to have called seqInit before calling any seqFoo% functions.% (including this one)if (nargin < 1)help seqExampleserror(`Which example?`)endswitch which%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% Simple example to get and view sequencescase 1disp(`Get sequence number 9`);seq= seqGet(9);disp(`View it`);seqView(seq);pause(2);disp(`Get sequence number 4`);seq = seqGet(4);disp(`View part of it (frames 20-40)`);seqView(seq, 20:40);pause(2);disp(`Get it from a different angle (pi/2)`);seq = seqGet(4, pi/2);disp(`View part of it (frames 20-40)`);seqView(seq, 20:40);pause(2);disp(`Get a sequence in 3d (default)`);seq3d = seqGet(6, `3d`, pi/2);disp(`Convert it to 2d`);seq2d = seqConvert(seq3d, `2d`); % Note: "seq2d = seqGet(6, '2d, pi/2)" is equiv to theabove . . .disp(`View the 2d sequence, highlighting leds #4, #10 (inred)`);seqView(seq2d, [ ], `leds`, [4 10])pause(2);disp(`Get the sequence as an image`);seqIm = seqGet(9, `image`);disp(`And view it.`);seqView (seqIm)pause(2);disp(`Get the sequence in 3d`);seq3d = seqGet(9);disp(`View them overlaid`);seqView2({seqIm, seq3d})pause(2);disp(`Get the sequence as form data`);seqFo = seqGet(9, `form`, pi/2);disp(`and view it.`);seqView(seqFo);%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% Training and Reconstruction of 3d and 2d datacase 2disp(`Select out all but sequence #6 at 90 degrees`);sel = seqselect([1:5 7:10], pi/2);disp(`Train on these sequences in 2d`);knl = seqTrain(sel, `2d`);disp(`Get an observation (in the same representation as training!(2d))`);obs = seqGet(6, `2d`, pi/2);disp(`Reconstruct that observation (into 3d with 100 e-vects)`);rcn = seqRecon(obs, knl, `3d`, 100);disp(`View the reconstructed sequence (in 3d)`);seqView(rcn)%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% Rendering 3d Data%% WARNING: Only call this once you've set the Vtk Display using%% seqSetVtkDisplay.case 3disp(`Get sequence #9`);seq = seqGet(9);disp(`Crop out the first 5 frames`);seq = seqCrop(seq, 5);disp(`Render at pi/4 degrees`);ren = seqRender(seq, pi/4);disp(`And view the result`);seqview(ren);%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% Image processing on New Data%% WARNING: This example is VERY SLOW!%%%% ANOTHER WARNING: I HAVENT HAD TIME TO TEST THIS **case 4disp(`Import a new sequence of pgm images`);seqIm = seqImport(`/homes/leventon/Images/Bill/wave.sub.-- %%05d.pgm',0:2:224);disp(`View it`);seqView(seqIm);disp(`Do foreground computation (SLOW!)`);seqFg = seqForeground(seqIm);disp(`View the first few frames`);seqView(seqFg, 1: 10); disp(`Downsample the result`);seqSm = seqDownsample(seqFg, 2); disp(`Convert the downsampled foreground image to formrepresentation`);seqFo = seqconvert(seqSm, `form`); disp(`Edge detect the regular-sized foreground image`);seqEd = seqEdge(seqFg);disp(`View the first few frames`);seqView(seqEd, 1:10);%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% Image processing on Bill's Datacase 5global seqImglobal seqFggiobal seqSmglobal seqFoglobal seqEdglobal rcnFoglobal bFoglobal ptcorrglobal trackcorrif (0)seqIm = seqImport(`/homes/leventon/Images/Bill/wave.sub.-- %05d.pgm`,0:2:224)seqFg = seqForeground(seqIm);seqsm = seqDownsample(seqFg, 2);seqFo = seqconvert(seqsm, `form`);seqEd = seqEdge(seqFg);elsedisp(`Loading Bills Sequence Data`);load Data/billwave.matendseqFoc = seqCrop(seqFo, 27);if (0)% starting with no correspondences[rcnFo, bFo] = seqRecon(seqFoc, knlFo, `3d`, 100, `reg`, 1e4, 1,0, 1, 1e8, `track`, seqEd, `interactive`, 1, seqIm);else% starting with the correspondences I set.disp(`Load Trainign Data`);load Data/knlFo-000-2.matdisp(`Loading Bills Tracking Data`);load Data/billtrack.matdisp (`Reconstructing`);[rcnFo, bFo, ptcorr, trackcorr] = seqRecon(seqFoc, knlFo, `3d`,100, `reg`, 1e4, 1, 0, 1, 1e8, `track`, seqEd, `interactive`, 1,seqIm, `ptcorr`, ptcorr, `trackcorr`, trackcorr);end%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% Image processing on Barish's Data with new training data%% See the file Fit-Example.trans to see a transcript.case 6global seqImglobal seqFgglobal seqSmglobal seqFoglobal seqEdglobal rcnFoglobal bFoglobal ptcorrglobal trackcorrdisp(`Now, calling seqGet with no args prints out a list of`);disp(`3d sequences that can be loaded:`);seqGetdisp(`You should be able to use all the new data in the same`);disp(`way that we used the old data, just by doing seqGet(53),etc . . .`);disp(`.`)disp(`Ive got the following image sequences saved in a .matformat.`);disp(`Assuming you want to use one of these, just load theappropriate`);disp(`.mat file from .sup..about. leventon/Data. The sequences I haveare`);disp(` barish.mat, billwave.mat, fred.mat, seqme.mat, trav.ma,`);disp(` wiz.mat, wiz2.mat`);disp(`.`);disp(`Loading barishs image data (please wait)`);load Data/barish.matdisp(`.`);whosdisp(`seqEd -- edge-dectected image`);disp(`seqIm -- Full size (320x240 image)`);disp(`seqsm -- Smaller image (80x60 image)`);disp(`.`);disp(`Below, find the best fit of the image sequence (seqSm) `);disp(`to subtype 2 (new data). Train on sequence #60 at 0degrees.`);disp(`(You could also do 60:65, [0 pi/2 to train on sequences60-65`);disp(` on both 0 and pi/2 rotations). Set interactive to 1,allowing`);disp(`the user to specify constraints. Supply a high-res imageseq`);disp(`seqIm for better viewing of the image when addingcorrespondences.`);disp(`Also, Im setting the ISSD (Intensity SumSqDiff) energy to 0`);disp(`for speed purposes during this first interactive phase.`);disp(`Im also setting the fminiter to 1000 (max # iterations)`);disp(`When the interactive window appears, click on a handful `);disp(`od corresp. in the first frame and then click "Done"`);[seq, alpha, beta, ptcorr, trackcorr, knl3d] = . . .seqFit(seqSm, `seqIm`, seqIm, `interactive`, 1, . . . `subtype`, 2, `train`, 60, 0, . . .`ISSDELambda`, 0, `fminiter`, 1000);disp(`Ok, lets look at what seqFit returned: `);disp(`[seq, alpha, beta, ptcorr, trackcorr, knl3d] = seqFit(. . .)`);disp(`seq == the reconstruction sequence`);disp(`alpha == matrix nEv x nSnip. Each colm vector is the set ofe-vect coeffs`);disp(`beta == matrix n x nSnip. Each colm vector is the set oftraining coeffs`);disp(` where N is the number of training snippets . . .`);disp(` and nSnip id the number of snippets in the inputimage seq.`);disp(`ptcorr == the list of pt correspondences that I clicked`);disp(`trackcorr == the list of tracking correspondences (null inthis example)`);disp(`knl3d == the knowledge base computed from the trainingdata.`);disp(`.`);disp(`Note that now, I can rerun seqFit non-interactively using `);disp(`the correspondences from last time (with different fmin lambdaparams)');disp(`Also I replaced "train" with "knl3d", since we already have the`);disp(`training data.`);disp(`I also passed in "alpha" (the coeffs) as a new initializationvector. . .`);disp(`.`);disp(`Finally, I passed in a parameter "snips" which tells it whichsnippets`);disp(`compute on (in the interactive mode, it only computessnippets');disp(`that you edit . . . Here, it would try to compute all 19snippets`);disp(`otherwise -- slow!). I told it just to compute snippet 1.`);disp(`You could pass in ( . . ., "snip", 1:4) to compute snips 1through 4, etc.`);[seq, alpha, beta] = . . . seqFit(seqSm, `seqIm`, seqIm, `interactive`, 0, . . . `subtype`, 2, `knl3d`, knl3d, . . .`ISSDELambda`, 0.05, `fminiter`, 500, . . . `alpha`, alpha, `snips`, 1);%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%otherwiseerror(`That example isnt available`);endfile: /homes/leventomn/Matlab/seqFit.mfunction [seq. alpha, beta, ptcorr, trackcorr, knkl3d] = seqFit(seqSm,varargin% seqFit Fit the 3D representation to an image sequence%% seqFit(seqIm, . . . options . . .);%% Options:% The initial sequence can be supplied as either a sequence% (using `seq`), a set of eigen-coefficients (using `alpha`),% or a set of training-snippet-coefficients (using `beta`).% If none are supplied, then the initial sequence is taken% from a training sequence of the appropriate subtype.% `seq`, seq% `alpha`, alpha% `beta`, beta% If the sequence is not supplied, then the subtype must% be given (otherwise it is assumed to be 1)% `subtype`, subtype% One of the following two must be supplied. knl3d i sthe% knowledge% based returned from a call to seqTrain. Otherwise, whichseq% and whichchang are used to call seqTrain. - % `knl3d`, knl3d% `train`, whichseq, whichang% Snips is an array of snippet numbers to operate on (i.e. 1:6)% `snips`, snips% seqIm is the full-resolution image sequence to use to display% `seqIm`, seqIm% If you want to track using Hausdorff distance, supply an edge% sequence% `track`, seqEdge% Starting point correspondences% `ptcorr`, ptcorr% Starting tracking correspondences% `trackcorr`, trackcorr% Interactive (1 or 0) if you want to allow the user to click% points% `interactive`, interactive% Number of e-values to use% `nEv`, nEv% Verbose messages (1 or 0)% `verbose`, verbose% Lambdas for minimization% `corrLambda`, corrLambda% `LenDELambda`, LenDELambda% `priorLambda`, priorLambda% `ISSDELambda`, ISSDELambda% Max number of minization iterations.% `fminiter`, fminiter%% DefaultsseqIm = seqSm;interactive = 0;ptcorr = { };trackcorr = { };seqEdge = [ ];nEv = 50;knl3d = [ ];whichseq = 1:10;whichang = pi/2;seq = [ ];alpha = [ ];beta = [ ];corrLambda = 0.5;priorLambda = 0.1;%% LenDELambda = 0.1;LenDELambda = 0;ISSDELambda = 0.05;verbose = 1;snips = [ ];fminiter = 500;subtype = 1;%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% Process Optionsargc = 1;while (argc <= nargin-1)if (isstr(varargin{argc}) == 0)disp(`Usage Error: Option is not a string`);help seqFiterror(`Usage Error: Uption is not a string`);endswitch varargin{argc}case `verbose`verbose = varargin{argc+1};argc = argc + 2;case `corrLambda`corrLambda = varargin{argc+1};argc = argc + 2;case `LenDELambda`LenDELambda = varargin{argc+1};argc = argc + 2;case `priorLambda`priorLambda = varargin{argc+1};argc = argc + 2;case `ISSDELambda`ISSDELambda = varargin{argc+1};argc = argc + 2;case `fminiter`fminiter = varargin{argc+1};argc = argc + 2;case `seq`seq = varargin{argc+1};argc = argc + 2;case `alpha`alpha = varargin{argc+1};argc = argc + 2;case `beta`beta = varargin{argc+1};argc = argc + 2;case `knl3d`knl3d = varargin{argc+1};argc = argc + 2;case `snips`snips = varargin{argc+1};argc = argc + 2;case `train`whichseq = varargin{argc+1};whichang = varargin{argc+2};argc = argc + 3;case `seqIm`%% Include a high resolution imageseqIm = varargin{argc+1};argc = argc + 2;case `track`seqEdge = varargin{argc+1},argc = argc + 2;case `ptcorr`ptcorr = varargin{argc+1};argc = argc + 2;case `trackcorr`trackcorr = varargin{argc+1};argc = argc + 2;case `interactive`interactive = varargin{argc+1};argc = argc + 2;case `nEv`nEv = varargin{argc+1};argc = argc + 2;case `subtype`subtype = varargin{argc+1};argc = argc + 2;otherwisedisp(`Usage Error: Bad Option`);help seqReconerror(`Usage Error: Bad Option`);endendseqIm.subtype = subtype;seqSm.subtype = subtype;if isempty(knl3d)if (verbose) disp(`Training . . .`); endknl3d = seqTrain(seqSelect(whichseq, whichang), `3d`, . . .`MtM`, 0, `submean`, 1, `headalign`, 1);endencParams = knl3d.encParams;Mnewrep = knl3d.enc{1};global seqINFO%%% HACK!nmk = seqINFO.nummkr(encParams.subtype);k = encParams.step;s = encParams.fps;numSnip = size(seqEncode(seqCrop(seqGet(10), seqSm.nframes), `force`,1) , 2)init = cell(numSnip,1);for i=1:numSnip; init{i}=cell(0); endif (isempty(snips)), snips = 1:numSnip; endS3i = spdiags(1./diag(knl3d.S), 0, size(knl3d.V,1), nEv);V3S3i = knl3d.V * S3i;if (isempty(alpha) .vertline. (size(alpha, 1) < nEv))if (isempty(beta))if (isempty(seq))if (subtype == 1) seq = seqCrop(seqGet(10), 1, seqSm.nframes);elseif (subtype == 2) seq = seqCrop(seqGet(51), 1, seqSm.nframes);endseq.scale = seqSm.scale;enc = seqEncode(seq, `force`, 1);beta = knl3d.V * S3i(:,1:nEv) * knl3d.U(:,i:nEv)' * enc;alpha = knl3d.S(1:nEv,:) * knl3d.V' * beta;else%% Seq exists. Encode, and compute alpha.enc = seqEncode(seq, `force`, 1);beta = knl3d.V * S3i(:,1:nEv) * knl3d.U(:,1:nEv)' * enc;alpha = knl3d.S(1:nEv,:) * knl3d.V' * beta;endelse%% Beta exists. Compute alpha from beta.alpha = knl3d.S(1:nEv,:) * knl3d.V' * beta;endendalpha = alpha(1:nEv, :);%% At this point, we have a valid alpha.beta = V3S3i * alpha;Mrecon = Mnewrep * beta;if (exists(knl3d, `mean`))Mrecon = Mrecon + knl3d.mean{1} * ones(1, size(Mrecon, 2));endseq = seqDecode(Mrecon, `3d`, encParams, seqSm);foptions(1) = 1;foptions(2) = 0.001;foptions(3) = 0.001;foptions(14) = fminiter;% foptions(16) = 0.00001;itering = 1;last.sub.-- t = 1;itercount = 1;if (isempty(seqEdge) .sup..about. = 1)edgeScale = seqEdge.scale / seqSm.scale;endnFr = seqSm.nframes;times.sub.-- changed = ones(1,seqsm.nframes);while (itering)%% Reset all the data matricesMtM = zeros(size(knl3d.S, 2));Mty = zeros(size(knl3d.S, 2), numSnip);M.sub.-- s = init;y.sub.-- s = init;weight.sub.-- s = cell(numSnip, 1);%% Interactive Editingif (interactive == 1)[ptcorr, trackcorr, itering, last.sub.-- t, times.sub.-- changed] = . ..addconstraints(seq, seqIm, ptcorr, trackcorr, last.sub.-- t);disp(`Saving ptcorr . . . Done`);save ptcorr.mat ptcorrif (itercount == 1), itering = 1; endif (itering == 0), break; endelseitering = 0;end%% Add the correspondence constraints to the observation[MtM, Mty, M.sub.-- s, y.sub.-- s, weight.sub.-- s] = compConstr(MtM,Mty, 0, . . .ptcorr, 1, seqSm, Mnewrep, nmk, encparams, numSnip, . . .M.sub.-- s, y.sub.-- s, weight.sub.-- s, knl3d.mean{1});if (isempty(seqEdge) .sup..about. = 1)%% Track based on the initial constraints[trackcorr, Models, Range, HausFrac, Certainty] = . . .seqTrack(seqEd, trackcorr, edgeScale);addconstraints(seq, seqIm, ptcorr, trackcorr, last.sub.-- t);%% Add the tracking constraints to the observation[MtM, Mty, M.sub.-- s, y.sub.-- s, weight.sub.-- s] = compConstr(MtM,Mty, 1,trackcorr, . . .1, seqSm, Mnewrep, nmk, encParams, numSnip, . . .M.sub.-- s, y.sub.-- s, weight.sub.-- s, knl3d.mean{1});endsnips.sub.-- changed = [ ];times.sub.-- changed = [times.sub.-- changed zeros(1,nFr)];for snip = snipsif (sum(times.sub.-- changed((k*(snip-1)+1):(k*(snip-1)+s))))snips.sub.-- changed = [snips.sub.-- changed snip];endendfor snip = snips.sub.-- changedseqSmCrop = seqCrop(seqSm, k*(snip-1)+1, k*(snip-1)+s);fEnergyInit2(`PriorE`, priorLambda, {S3i}, . . . `ISSDE`, ISSDELambda, {Mnewrep, V3S3i, seqSmCrop, 1,knl3d.mean{1}, encParams, 0.5}, . . .`LenDE`, LenDELambda, {Mnewrep, V3S3i, 1}, . . . `CorrE`, corrLambda, {M.sub.-- s(snip), y.sub.-- s(snip), . . .weight.sub.-- s(snip), MtM(snip), Mty(:,snip), nEv, V3S3i});alpha(:,snip) = fminu(`fEnergy2`, alpha(:,snip), foptions);beta(:,snip) = V3S3i * alpha(:,snip);end%% ProjectingMrecon = Mnewrep * beta;if (exists(knl3d, `mean`))Mrecon = Mrecon + knl3d.mean{1} * ones(1, size(Mrecon, 2));end%% Decodeseq = seqDecode(Mrecon, `3d`, encParams, seqSm);itercount = itercount + 1;endfile: /homes/leventon/Matlab/seqForeground.mfunction [seqIm, bgimage] = seqForeground(seqIm, bgimage, bgmotion)% seqForeground Compute the Foreground of an Image Sequence%% [seqIm, bgimage] = seqForeground(seqIm) computes the mode% background image% and then segments out the foreground from the background%% [seqIm, bgimage] = seqForeground(seqIm, bgimage) uses the bgimage% to segment out the foreground from the background.%% [seqIm, bgimage] = seqForeground(seqIm, [ ], 1) first computes the% best% affine transform between every frame and the middle frame,% and then computes the mode background image% and then segments out the foreground from the background.%mode.sub.-- width = 8;if (nargin < 2)bgimage = [ ];endif (nargin < 3)bgmotion = 0;endnfr = seqIm.nframes;for i = 1:nfrI(:,:,i) = seqIm.data{i};end[m, n, lose] = size(I);if (bgmotion)disp(`seqForeground: Factoring out image motion (slow!)`);reffn = sprintf(`tmp-%04.ref`, round(mod(now, 10000)));tmpfn = sprintf(`tmp-%04.tmp`, round(mod(now, 10000)));outfn = sprintf(`tmp-%04.out`, round(mod(now, 10000)));infn = sprintf(`tmp-%04d.in`, round(mod(now, 10000)));ref = floor((nfr+1)/2);I = I + min(I(:));mI = max(I(:));if (mI .sup..about. = 0), I = I * 255 / mI; endfid = fopen(reffn, `w`);fwrite(fid, I(:,:,ref), `uchar`);fclose(fid);w = size(I,1);h = size(I,2);for i = [1:ref-1 ref+1:nfr]fid = fopen(tmpfn, `w`);fwrite(fid, I(:,:,i), `uchar`);fclose(fid);cmd = sprintf(`cat %s %s .vertline./homes/leventon/DYNAMO/src/shell-src/affine-estimate-level 4intype ib -xdim %d -ydim %d -start 0 -track -mask "r: %d %d %d%d" >! %s', reffn, tmpfn, w, h, 10, 10, w-10, h-10, outfn)unix(cmd);cmd = sprintf(`cat %s .vertline./homes/leventon/DYNAMO/src/shell-src/warp-image -intype ib -xdim%d -ydim %d -rxdim %od -rydim %d -paramfile %s -n 1 -zoom 1 -xzoom1 -yzoom 1 -adjust "0 0 0 0 0 0"-invert -outtype ib > ! %s',tmpfn, w, h, w, h, outfn, infn)unix(cmd);fid = fopen(infn, `r`);I(:,:,i) = reshape(fread(fid, w*h, `uchar`), w, h);fclose(fid);cmd = sprintf(`/bin/rm -f %s %s %s`, tmpfn, outfn, infn);unix(cmd);endendif (0)for i = 1:nfrseqIm.data{i} = I(:,;,i);endreturnendif (isempty(bgimage))disp(`seqForeground: Finding the Mode Background Image (a littleslow) `)bgimage = mode(I, 3, mode.sub.-- width);endbgimage = bgimage - median(bgimage(:));disp(`seqForeground: Segmenting person (slow! (Get a drink))`);for i = 1:nfrON.sub.-- FRAME = iIm = I(:,:,i) - median(vectorize(I(:,:,i)));Id = imDiff(Im, bgimage, 10);Id = Id / max(max(Id));Ithresh = Id>0.10;Ierode = erode(Ithresh, ones(3));Idilate = double(dilate(Ierode, circle(6)));Ismooth(:,:,i) = smooth(Idilate,8);endFG = reshape(Ismooth, m*n, nfr);FG = conv2(1, [1 4 6 4 1], FG, `same`);FG = reshape(FG, m, n, nfr);FG = FG .* I;for i = 1:nfrseqIm.data{i} = FG(:,:,i);endfile: /homes/leventon/Matlab/seqGet.mfunction seq = seqGet(n, rep, varargin)% seqGet Get a Sequence from the sequence database%% seq = seqGet(n, rep, theta, texture) returns the nth sequence of% the% database in representation `rep`, with rotation theta.%% The texture is not currently used (but allowed adding texture to% the% Vtk rendered figures)%% Defaults: rep = `3d`, theta = 0, texture = 0%% Valid representations include:% `3d`, `2d`, `ang`, `image`, `motion`, `form`, `stick`, `edge`,% `head`% `pulse`%global seqINFOif (nargin < 1)%% Don't get anything . . . Print Descriptionsfor i = 1:length(seqINFO.seqs.names)if (isempty(seqINFO.seqs.names{i})==0)disp(sprintf(`# %3d is %-10s.backslash.%4d frames)`, i, . . .seqINFO.seqs.names{i}, seqINF0.seqs.nframes(i)));endendreturnendif (nargin < 2), rep = `3d`; endparams = varargin;if (length(params) < 1)theta = seqINFO.defaults.orientation;elsetheta = params{1};endif (length(params) < 2)texture = seqINFO.defaults.texture;elsetexture = params{2};endif (isstr(rep) == 0)if (isstr(theta) == 1)tmp = rep;rep = theta;theta = tmp;params{1} = theta;elsetheta = rep;rep = `3d`;endendif (seqINFO.defaults.debug)disp(sprintf(`Getting seq #%d in rep %s at angle %d . . .`, n, rep,. . .round(theta*180/pi)));endswitch repcase `3d`if (isempty(seqINFO.rep3d. data{n}))fn = sprintf(seqINFO.rep3d.filespec, n);load(fn);newdata = seq;if (n > 10)newdata = newdata/30;newdata(1,:) = seq(1,:);endclear seqseq.type = `3d`;seq.subtype = seqINFO.seqs.subtypes(n);seq.name = seqINFO.seqs.names{n};seq.data = newdata;seq.angle = 0;seq.nframes = newdata(1,1);seq.head = seqGet(n, `head`, theta);if (seqINFO.defaults.cache3d) seqINFO.rep3d.data{n} = seq;endelseseq = seqINFO.rep3d.data{n};endseq = seqRotate(seq, theta);case ` 2d`seq = seqconvert(seqGet(n, `3d`, params{1:length(params)}), `2d`);case `image`nStored length(seqINFO.repImage.data{n});found = 0;for i=1:nStoredif (seqINFo.repImage.data{n}{i}.angle == theta)if (seqINFO.repImage.data{n}{i}.texture == texture) found = 1; seq = seqINFO.repImage.data{n}{i}; endendendif (found == 0)s = sprintf(seqINFO.repImage.filespec, n, round(theta/pi*180));fprintf(1, `seqGet: Reading image sequence: %s.backslash.n`, s);load(s);for i = 1:seq.nframesseq.data{i} = double(seq.data{i});endif (seqINFO.defaults cacheImage) seqINFO.repImage.data{n}{nStored+1} = seq;endendcase `motion`nStored = length(seqINFO.repMotion.data{n});found = 0;for i=1:nStoredif (seqINFO.repMotion.data{n}{i}.angle == theta)if (seqINFO,repMotion.data{n}{i}.texture == texture) found = 1; seq = seqINFO.repMotion.data{n}{i}; endendendif (found == 0)s = sprintf(seqINFO.repMotion.filespec, n, round(theta/pi*180));if (exist(s) == 2) fprintf(i, `seqGet: Loading motion sequence: %s.backslash.n`, s); load(s); for i = i:seq.nframes seq.data{i} = double(seq.data{i})/255*(2*pi) - pi; end returnendseq = seqConvert(seqGet(n, `image`, params{1:length(params)}), `motion`);seq.nframes = length(seq.data);if (seqINFO.defaults. cacheMotion) seqINFO.repMotion.data{n}{nStored+1} = seq;endendcase `form`nStored = length(seqINFO.repForm.data{n});found = 0;for i=1:nStoredif (seqINFO.repForm.data{n}{i}.angle == theta)if (seqINFO.repForm.data{n}{i}.texture == texture) found = 1; seq = seqINFO.repForm.data{n}{i}; endendendif (found == 0)s = sptintf(seqINFO.repForm.filespec, n, round(theta/pi*180));if (exist(s) == 2) fprintf(1, `seqGet: Loading form sequence: %s.backslash.n`, s); load(s); returnendseq = seqConvert(seqGet(n, `image`, params{1:length(params)}),`form`);seq.nframes = length(seq.data);if (seqINFO.defaults.cacheForm) seqINFO.repForm.data{n}{nstored+1} = seq;endendcase `pulse`nStored = length(seqINFO.repPulse.data{n});found = 0;for i=1:nStoredif (seqINFo.repPulse.data{n}{i}.angle == theta) found = 1; seq = seqINFO.repPulse.data{n}{i};endendif (found == 0)s = sprintf(seqINFO.repPulse.filespec, n, round(theta/pi*180));if (exist(s) == 2) fprintf(1, `seqGet: Loading pulse sequence: %0s.backslash.n`, s); load(s); if (seqINFO.defaults.cachePulse) seqINFO.repPulse.data{n}{nStored+1} = seq; end returnendseq = seqZConvert(seqGet(n, `2d`, params{1:length(params)}),`pulse`);seq.nframes = length(seq.data);if (seqINFO.defaults.cachePulse) seqINFO.repPulse.data{n}{nstored+1} = seq;endendcase `stick`nStored = length(seqlNFO.repStick.data{n});found = 0;for i=1:nStoredif (seqINFO.repstick.data{n}{i}.angle == theta)if(seqINFO.repstick.data{n}{i}.texture == texture) found = 1; seq = seqINFO.repStick.data{n}{i}; endendendif (found == 0)s = sprintf(seqINFO.repStick.filespec, n, round(theta/pi*iBO));if (exist(s) == 2) fprintf(1, `seqGet: Loading stick sequence: %s.backslash.n`, s); load(s); returnendseq = seqConvert(seqGet(n, `3d`, params{1:length(params)}),`stick`)seq.nframes = length(seq.data);if (seqINFO.defaults.cacheStick) seqINFO.repStick.data{n}{nStored+1} = seq;endendcase `edge`nStored = length(seqINFO.repEdge.data{n});found = 0;for i=1:nStoredif (seqINFO.repEdge.data{n}{i}.angle == theta)if (seqINFO.repEdge.data{n}{i}.texture == texture) found = 1; seq = seqINFO.repEdge.data{n}{i}; endendendif (found == 0)s = sprintf(seqINFo.repEdge.filespec, n, round(theta/pi*i80));if (exist(s) == 2) fprintf(1, `seqGet: Loading edge sequence: %s.backslash.n`, s); load(s); returnendseq = seqConvert(seqGet(n, `3d`, params{1:length(params)}),`edge`);seq.nframes = length(seq.data);if (seqINFO.defaults.cacheEdge) seqINFO.repEdge.data{n}{nStored+1} = seq;endendcase `head`nStored = length(seqINFO.repHead.data{n});found = 0;for i=1:nStoredif (seqINFO.repHead.data{n}{i}.angle == theta) found = 1; seq = seqINFO.repHead.data{n}{i};endendif (found == 0)fn = sprintf(seqINFO.repHead.filespec, n);load(fn);if (n > 10) head = head / 30;endseq.data3d(:,1) = cos(theta) * head(:,1) + -sin(theta) *head(:,2);seq.data3d(:,2) = sin(theta) * head(:,1) + cos(theta) *head(: ,2);seq.data3d(:,3) = head(:,3);seq.data2d = worldToCam(seq.data3d, 0);seq.type = `head`;seq.subtype = seqINFO.seqs.subtypes(n);seq.name = seqINFO.seqs.names{n};seq.angle = theta;seq.nframes = size(head,1);clear headif (seqINFO.defaults.cacheHead) seqINFO.repHead.data{n}{nStored+1} = seq;endendcase `ang`seq = seqConvert(seqGet(n, `3d`, params{1:length(params)}),`ang`);otherwiseerror(`seqGet: Bad Representation`);endif (seqINFO.defaults.debug)disp(sprintf(`Done`));endfile: /homes/leventon/Matlab/seqHelp.m%--------------------------------------------------------file: %Human Motion Tracking Project Prototype for Matlab% Michael Leventon MERL Summer 1997% leventon#ai.mit.edu%----------------------------------------------------------file: %Sequence Abstraction%% seqInit - Initialize all Sequence Functions% seqExamples - Examples how to use the abstraction%% seqGet - Get a sequence from sequence database% seqConvert - Convert a sequence frorn one representation to% another% seqCrop - Crop a sequence (in time) or select out frames.% seqRotate - Rotate a sequence%% seqSelect - Select which sequences and directions to train from% seqTrain - Train on a set of sequences% seqRecon - Reconstruct an observation into a new representation%%% seqView - View a sequence% seqView2 - View multiple sequences together (ALPHA!)%% seqData - Accessor function returning the raw sequence data%% seqRender - Render a 3d sequence as a cyiinder model with VTk% seqSetVtkDisplay - Set the display for Vtk to use in seqRender%% seqImport - Import a bunch of pgm frames as an image sequence% seqForeground - Segment the foreground (person) in an image% sequence% seqDownsample - Downsample an image sequence%% seqEdge - Canny Edge Detect an image sequence% seqTrack - Track parts of an edge sequence over time%% seqPaste - Paste two image sequences side by side (for% visualization)% seqImSave - Save an image sequence as a series of pgm frames%% seqDecode - Decode raw data into a sequence (used internally)% seqEncode - Encode a sequence as a matrix of snippets (used% internally)%% seqError - Compute the 2d error between two sequences% seqKE - Compute the abs change in Kinetic Energy of a 3d% sequence% seqLengthE - Compute the length energy of a 3d sequence% seqSmoothE - Compute the smoothness energy of a 3d sequence% seqTorqueE - Compute the torque energy of a 3d sequencehelp seqHelpfile: /homes/leventon/Matlab/seqImSave.mfunction seqImSave(seq, filespec, frames)if (strcmp(seq.type, `image`) == 0)error(`Can only write out image sequences`);endnfr = seq.nframes;if (nargin < 3), frames = 1:nfr; endminval = seq.data{1}(1,1);maxval = seq.data{1}(1,1);for i = framesminval = min(minval, min(seq.data{i}(:)));maxval = max(maxval, max(seq.data{i}(:)));endfor i = framess = sprintf(filespec, i);pgmWrite(seq.data{i}, s, [minval maxval]);endfile: /homes/leventon/Matlab/seqImport.mfunction seqIm = seqImport(fn, range, desc, subtype, pos)% seqImnport Import an image sequence frorn a series of pgm images%% seqIm = seqImport(fn, range, desc) reads in the sequence of pgm%% images. "fn" should have a %d (or %05d, etc) specifier.% Range is a vector of image numbers (1:10, 30:50, 1:2:100,%etc),% for the images to be read. Desc is a name for the sequence.% By default, desc is set to the file name.%% The user is asked to click on the head of the person in% each frame. We should be able to do something better than%this% (like track the head, since we're doing it anyway).if (nargin < 3), desc = [ ]; endif (isempty(desc)), desc = fn; endif (nargin < 4)disp(`Warning: Assuming subtype equals 1`);subtype = 1;endif (nargin < 5), pos = [ ]; endseqIm.type = `image`;seqIm.name = desc;seqlrn.subtype = subtype;seqIm.angle = 0;seqIm.texture = 0;seqIm.nframes = length(range);seqIm.scale = 1;disp(`With each new frame, please click on the head of the person`);disp(`(yeah, we could track it, but I am running out of time)`);count = 1;clfneedtoclick = isempty(pos);for i = ranges = sprintf(fn, i);seqIm.data{count} = pnmRead(s);if (size(seqIm.data{count}, 3) == 1)colormap (gray);imagesc(seqIm.data{count});else% rgbimagesc(seqIm.data{count}); imagesc(seqIm.data{count}(:,:,2));endaxis image;if (needtoclick)pos(count,:) = ginput(1);endcount = count + 1;endseqIm.head.data2d = pos;seqIm.head.type = `head`;seqIm.head.subtype = subtype;seqIm.head.name = desc;seqIm.head.angle = 0;seqIm.head.nframes = length(range);file: /homes/leventon/Matlab/seqInit.m%% Initialization file for the sequence functions%% This must be called before any other are calledclear seqINFOglobal seqINFOpath(path, `/homes/leventon/matlabpyrTools`);seqINFO.defaults.debug = 0;seqINFO.defaults.texture = 0;seqINFO.defaults.orientation = 0;seqINFO.defaults.snipsize = 10;seqINFO.defaults.snipstep = 5;seqINFO.defaults.rendercmd = `/homes/leventon/Vtk/renderguy.tcl`;seqINFO.defaults.model = `homes/leventon/Matlab/Temp/person.mdl`;seqINFO.defaults.seq = `/homes/leventon/Matlab/Temp/seq.seq`;seqINFO.defaults.framespec =`/homes/leventon/Matlab/Temp/tmp-%03d ppm`;seqINFO.defaults.display = ``;seqINFO.defaults.cacheImage = 0;seqINFO.defaults.cacheMotion = 0;seqINFO.defaults.cacheForm = 0;seqINFO.defaults.cache3d = 0;seqINFO.defaults.cacheHead = 0;seqINFO.defaults.cacheAngle = 0;seqINFO.defaults.cachestick = 0;seqINFO.defaults.cachePulse = 0;%% Subscripted by subtypeseqINFO.nummkr(1) = 37;seqINFO.connections{1} = [0 5; 5 6; 0 8; 8 9; 0 1; 0 4; 0 7; 4 7; 724; 24 25; 25 26; 9 25; 6 22; 4 21; 21 22; 22 23; 10 11; 11 12; 12 13;13 31; 27 28; 28 29; 29 31; 30 31; 32 33; 14 15; 15 16; 16 17; 17 36;33 34; 34 36; 35 36; 3 19; 19 20; 20 3; 7 15; 15 33; 33 7; 4 28; 2811; 11 4; 1 15; 1 11; 17 35; 13 30]seqINFO.connections{1} = [0 5; 5 6; 0 8; 8 9; 0 4; 0 7; 4 7; 7 24; 2425; 25 26; 9 25; 6 22; 4 21; 21 22; 22 23; 10 11; 11 12; 12 13; 13 31;27 28; 28 29; 29 31; 30 31; 32 33; 14 15; 15 16; 16 17; 17 36; 33 34;34 36; 35 36; 3 19; 19 20; 20 3; 7 15; 15 33; 33 7; 4 28; 28 11; 11 4;17 35; 13 30];seqINFO.nummkr(2) = 20;seqINFO.connections{2} = [1,2; 1,3; 2,3; 3,4; 3,5; 4,5; 4,6; 5,7; 6,8;. . .7,9; 8,10; . . . 9,11; 4,13; 5,14; 13,14; 13,12; 12,14; 13,15; 14,16;. . . 15,17; 16,18; 17,19; 18,20]-1;seqINFO.reps = {`3d`, `2d`, `image`, `motion`, `form`, `head`, `ang`,`pulse`};seqINFO.nseqs = 70;seqINFO.seqs.subtypes(1:10) = 1;seqINFO.seqs.subtypes(51:67) = 2;seqINFO.seqs.nframes(1:10) = [222 219 209 127 288 89 26 142 22 287];seqINFO.seqs.nframes(51:67) = [600 1500 2700 1479 600 2760 1620 1200600 1400 1500 1500 1500 2400 2700 2700 2700];seqINFO.seqs.names(1:10) = {`dance01`, `dance02`, `dance03`,`dance04`, `destroy`, `walkloop`, `runloop`, `fwalk`, `frun``thief`};seqINFO.seqs.names(51:67) = {`BillCal`, `Bill20`, `Bill45`, `BillImp`,. . .`EgonCal`, `Egon45a`, `Egon45b`, `EgonImp`, `MarthaCal`, . . .`MarthaBal20`, `MarthaLLBal20`, `MarthaModern20`, . . .`MarthaRobot20`, `MarthaBal45`, `MarthaLBal45`,`MarthaLLBal45`,`MarthaRobot45`};%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% 3d Representation%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%seqINFO.rep3d.name = `3d`;seqINFO.rep3d.filespec = `/homes/leventon/Data/Pts/seq.%02d`;seqINFO.rep3d.data = cell(seqINFO.nseqs,1);%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% 2d Representation%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%seqINFO.rep2d.name = `2d`;seqINFO.rep2d.data = cell(seqINFO.nseqs,1);%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% Image Representation%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%seqINFO.repImage.name = `image`;seqINFO.repImage.data = cell(seqINFO.nseqs,1);seqINFO.repImage.filespec =`/homes/leventon/Data/Images/seq-i-%02-%03d.mat`;%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% Motion Representation%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%seqINFO.repMotion.name = `motion`;seqINFO.repMotion.data = cell(seqINFO.nseqs, 1);seqINFO.repMotion.filespec =`/homes/leventon/Data/Motion/seq-m-%02d-%03d.mat`;%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% Form Representation%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%seqINFo.repForm.flame = `form`;seqINFO.repForm.data = cell(seqINFO.nseqs,1);seqINFO.repForm.filespec =`/homes/leventon/Data/Form/seq-%02d-%03d.mat`;%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% Pulse Representation%%%seqINFO.repPulse.name = `pulse`;seqINFO.repPulse.data = cell(seqINFO.nseqs,1);seqINFO.repPulse.filespec =`/homes/leventon/Data/Pulse/seq-%02d-%03d.mat`;/ - %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% Stick Representation%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%seqINFO.repStick.name = `stick`;seqINFO.repstick.data = cell(seqINFO.nseqs,1);seqINFO.repstick.filespec =`/homes/leventon/Data/Stick/seq-%02d-%03d.mat`;%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% Edge Representation%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%seqINFO.repEdge.name = `edge`;seqINFO.repEdge.data = cell(seqINFO.nseqs,1);seqINFO.repEdge.filespec =`/homes/leventon/Data/Edge/seq-%02d-%03d.mat`;%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% Head Representation%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%seqINFO.repHead.filespec = `/homes/leventon/Data/Head/head.%02d`;seqINFO.repHead.data = cell(seqINFO.nseqs, 1);%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% Angle Representation%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%tmp = seqGet(9);seqINFO.repAngle.model = mcylmodel(tmp.data);seqINFO.rep2d.data = cell(seqINFO.nseqs,1);clear tmp%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% Joint Length Data%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%dancecl = [0 5; 5 6; 0 8; 8 9; 0 1; 0 4; 0 7; 4 7; 7 24; 24 25; 25 26;9 25; 6 22; 4 21; 21 22; 22 23; 10 11; 11 12; 12 13; 13 31; 27 28; 2829; 29 31; 30 31; 32 33; 14 15; 15 16; 16 17; 17 36; 33 34; 34 36; 3536; 3 19; 19 20; 20 3; 7 15; 15 33; 33 7; 4 28; 28 11; 11 4; 1 15; 111; 17 35; 13 30];seqINFO.clengths = [2 3 18.1593; 2 11 15.9428; 2 15 15.9429;2 1920.3029; 3 4 5.3262; 3 11 3.1710; 3 15 3.1710; 3 19 9.0798; 4 203.9984; 4 21 7.5175; 5 6 18.4018; 6 7 17.6462; 6 22 4.4117; 6 2321.2940; 7 22 18.1893; 8 9 18.1921; 9 10 17.3450; 9 25 4.3363; 9 2621.2032; 10 25 17.8786; 11 12 5.6688; 11 15 2.1829; 11 19 7.7130; 1128 5.6685; 12 13 12.0170; 12 28 8.0166; 12 29 3.0044; 13 14 9.8775; 1329 12.3870; 13 30 2.4694; 13 31 17.0529; 13 32 13.6592; 14 30 10.1815;14 31 7.1763; 14 32 4.2993; 15 16 6.0651; 15 19 7.7131; 15 33 6.0653;16 17 11.8582; 16 33 8.5775; 16 34 2.9644; 17 34 12.2232; 17 3617.1757; 18 36 7.0667; 18 37 4.2336; 20 21 8.5148; 25 26 21.4875; 3031 17.2230; 30 32 13.4479; 31 32 4.2991; 35 36 17.3783; 36 37 4.2336];seqINFO.fixed.sub.-- connections = [2 3; 2 11; 2 15; 2 19; 3 4; 3 11; 315; 319; 4 20; 4 21; 5 6; 6 7; 6 22; 6 23; 7 22; 8 9; 9 10; 9 25; 9 26; 1025; 11 12; 11 15; 11 19; 11 28; 12 13; 12 28; 12 29; 13 14; 13 29; 1330; 13 31; 13 32; 14 30; 14 31; 14 32; 15 16; 15 19; 15 33; 16 17; 1633; 16 34; 17 34; 17 36; 18 36; 18 37; 20 21; 25 26; 30 31; 30 32; 3132; 35 36; 36 37];file: /homes/leventon/Matlab/seqKE.mfunction [e, ea] = seqKE(seq)d = seq.data;nfr = d(1,1);n = size(d,1);d = d(2:n, :);e = 0;ea = zeros(37,nfr-2);for t = 2:nfr-1b = [(t-2)*37 (t-1)*37 t*37];for i = 1:37p = [d(b(1)+i, :) ; d(b(2)+i, :); d(b(3)+i,:)];ei = ( normsq(p(3, :)-p(2, :)) - normsq(p(2, :)-p(1, :)) ). 2;ea(i, t-1) = ei;endende = mean(ea(:));file: /homes/leventon/Matlab/seqLeave1Out.mfunction errs = seqLeave1Out(sel, rep, nEv, dnsamp)if (nargin < 2)rep = `motion`;endif (nargin < 3)nEv = 50;endif (nargin < 4)dnsamp = 3;endencParams{1} = 10;encParams{2} = 5;encParams{3} = 37;encParams{4} = dnsamp;nums = sel.seq;ori = sel.ori;txtr = sel.texture;global M3dglobal MMo1global MMo3global MMo6global errsif (1)c = 1;for n = nums;for j = ori;for k = txtr;seqs{c} = seqGet(n, `3d`, j, k);c = c + 1;endendend[M3d, encParams] = seqEncode(seqs, encParams);endc = 1;for n = nums;for j = ori;for k = txtr;seqs{c} = seqGet(n, rep, j, k);c = c + 1;endendendif (1)encParams{4} = 1;[MMo1, encParams] = seqEncode(seqs, encParams);endif (strcmp(rep,`2d`) == 0)ENCODING = 3encparams{4} = 3;[MMo3, encParams] = seqEncode(seqs, encParams);ENCODING = 6encParams{4} = 6;[MMo6, encParams] = seqEncode(seqs, encParams);ENCODING = 0endseqs = 0;nSnip = size(M3d, 2);errs = zeros(4,nsnip);n = size(M3d, 1)for i = 1:5:nSnip[U, S, V] = svd(M3d(:, [1:(i-1) (i+1):nSnip]), 0);Si = spdiags(1./diag(S), 0, size(V,1), nEv);S = 0;if (nEv > size(U,2))nEv = size(U,2);fprintf(1, `nEv = %d.backslash.n`, nEv);endcoeff = (U(:,1:nEv)` * M3d(:, i));Mp3d = U(:, 1:nEv) * coeff;pi = [M3d(1:3:n, i) M3d(2:3:n, i) M3d(3:3:n, i)];pj = [Mp3d(1:3:n) Mp3d(2:3:n) Mp3d(3:3:n)];errs(1,i) = mean(vrnNorm(pi - pj, 2). 2);X = (V * (Si * coeff));if (strcmp(rep, `2d`))MpMo1 = MMo1(:, [1:(i-1) (i+1):nSnip]) * X;n2 = size(MMo1, 1)pi = [MMo1(1:2:n2, i) MMo1(2:2:n2, i)];pj = [MpMo1(1i:2:n2) MpMo1(2:2:n2)];errs(2,i) = mean(vmNorm(pi - pj, 2). 2);elseMpMo1 = MMo1(:, [1:(i-1) (i+1):nSnip]) * X;errs(2,i) = mean(normsq(MpMo1-MMo1(:,i)));MpMo3 = MMo3(:, [i:(i-1) (i+1):nSnip]) * X;errs(3,i) = mean(normsq(MpMo3-MMo3(:,i)));MpMo6 = MMo6(:, [i:(i-1) (i+1):nSnip]) * X;errs(4,i) = mean(normsq(MpMo6-MMo6(:,i)));endfprintf(1, `%d . . . ', nSnip-i);endfprintf(1, `.backslash.n`);file: /homes/leventon/Matlab/seqLengthE.mfunction e = seqLengthE(seq)global LENDATAd = seq.data;nfr = d(1,1);n = size(d,1);d = d(2:n, :);e = 0;len = LENDATA;for t = 1:nfrb = (t-1)*37;for i = 1:size(len,1)1 = norm((d(b+len(i,1),:) - d(b+len(i,2),:)));ei = normsq(1 - len(i,3));e = e + ei;endende = sqrt(e/t);file: /homes/leventon/Matlab/seqMkrData.mfunction data = seqMkrData(seq);data = seq.data(2:size(seq.data,1), :);data = reshape(data, seq.data(1,2), seq.data(1,1), size(seq.data, 2));data = permute(data, [1 3 2]);file: /homes/leventon/Matlab/seqNoise.mfunction seqn = seqNoise(seq, s)seqn = seq;if (s == 0)returnendswitch seq.typecase `2d`nfr = seq.data(1,1);n = nfr * 37;A2 = seq.data(2:n+1, :);for i = 1:ntheta = rand * 2 * pi - pi;v = [cos(theta) sin(theta)] * s * randn(1);A2(i, :) = A2(i, :) + v;endseqn.data(2:n+1, :) = A2;case {`3d`, `ang`, `image`, `motion`}error(`Not implemented`);endfile: /homes/leventon/Matlab/seqPairDiff.mfunction errs = seqpairDiff(sel, rep, dnsamp)if (nargin < 2)rep = `3d`;endnums = sel.seq;ori = sel.ori;txtr = sel.texture;c = 1;for n = nums;for j = ori;for k = txtr;seqs{c} = seqGet(n, rep, j, k);c = c + 1;endendendif (nargin < 3)dnsamp = 3;endencParams{1} 10;encParams{2} 5;encParams{3} 37;encParams{4} = dnsamp;[M, encParams] = seqEncode(seqs, encParams);nSnip = size(M, 2);errs = zeros(nSnip,nSnip);switch repcase `3d`n = size(M, 1);for i = 1:nSnipfprintf(1, `%d . . .`, nSnip-i);for j = i+1:nSnip pi = [M(1:3:n, i) M(2:3:n, i) M(3:3:n, i)]; pj = [M(1:3:n, j) M(2:3:n, j) M(3:3:n, j)]; errs(i,j) = mean(vmNorm(pi - pj, 2));endendcase `2d`for i = 1:nSnipfprintf(1, `%d . . `, nSnip-i);for j = i+1:nSnip pi = [M(1:2:n, i) M(2:2:n, i)]; pj = [M(1:2:n, j) M(2:2:n, j)]; errs(i,j) = mean(vmNorm(pi - pj, 2));endendcase `motion`for i = 1:nSnipfprintf(1, `%d . . .`, nSnip-i);for j = i+1:nSnip errs(i,j) = mean(abs(M(:, i) - M(:, j)));endendendfprintf(1, `.backslash.n`);file: /homes/leventon/Matlab/seqPaste.mfunction seq = seqPaste(seq1, seq2)% seqPaste Paste two image sequences together side by side% (for display purposes only)%% seq = seqpaste(seq1, seq2)%if ((strcmp(seq1.type, `image`) == 0) .vertline. (strcmp(seq2.type,`image`) ==0))error(`Both seqs must be of type image`);endnfr = min(seq1.nframes, seq2.nframes);h = min(size(seq1.data{1}, 1), size(seq2.data{1}, 1));w = min(size(seq1.data{1}, 2), size(seq2.data{1}, 2));seq = seq1;seq.data = cell(0);for i = 1:nfrseq.data{i} = [imselect(seq1.data{i}, [1 1; h w])imSelect(seq2.data{i}, [1 1; h w])];endseq.name = [`Pasted (`seq1.name `) and (`seq2.name `)`];seq.nframes = nfr;file: /homes/leventon/Matlab/seqRecon.mfunction [seq, betai, ptcorr, trackcorr, Mnewrep, coeffs] =seqRecon (varargin)% seqRecon sequence Reconstruction% seqRecon(obs, knl, newrep, nEv, . . . options . . .)%% seqRecon(obs1, . . ., obsN, knls, newrep, nEv, . . . options . . .)%% Options:% `core`, coreLambda, coreNumIter% `clip`, clipLambda, clipNumIter% `sumi`, sum1Lambda, sum1NumIter% `reg`, lambdas, nIter, core?, clip?, sum1% `weights`, weights (not used)% `Mnewrep`, Mnewrep% `interactive`, corrLambda, seqIm% `track`, seqEdge% `energymin`, lambdaKE, lambdaLenE% `pulse`, knlPulse, seqIm, pulse.sub.-- stacked.sub.-- rep, pulse.sub.--sensor.sub.-- pos,% . . .% pulse.sub.-- sensor.sub.-- ori, . . .% pulse.sub.-- twosigmasq, pulse.sub.-- cos.sub.-- power% `gausprior`, lambdaGausPrior% `trackcorr`, trackcorr% `ptcor`, ptcorr%% Return values: [seq, betai, ptcorr, trackcorr] = seqRecon(. . .)% seq -- the reconstructed sequence% betai -- the coefficients of the training data% ptcorr -- the point correpondences% trackcorr -- the tracking correpondencesif (nargin < 2)disp(`Usage Error 1`);help seqReconerror(`Usage Error`);endglobal MtM MtMO Mty MtyO M.sub.-- s y.sub.-- s weight.sub.-- s seqPulseencPulse obsLambdalamPulse pulse.sub.-- goal bet0 Mnewrep Mrecon betai coeffs Mobs bet%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% DEFAULTS% Reconstruct in 3dnewrep = `3d`;% Using 50 EigenvectorsnEv = 50;% Assume no regularization {lammbdas, nIter, core?, clip?, sum1}reg = {0 1 0 0 0};% Don't Tracktrack = 0;% Assume we have to encode the training data in the new representation%Mnewrep = [ ];% Assume we don't have any point correspondencesptcorr = cell(0);% Assume we don't have any tracking correspondencestrackcorr = cell(0);% Assume we don't want to energy minimizeenergymin = 0;% Default lambda for correlationscorrLambda = 1;% Assume we don't want interactiveinteractive = 0;% Assume we aren't using pulse minimizationpulse = 0; - % Assume no gaussian prior on the coefficientslambdaGausPrior = 0;betai.sub.-- given = [ ];%lamLenE = 0;lamKE = 0;%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% Read the observation sequence(s)argc = 1;while (argc <= nargin)if (isSequence(varargin{argc}) == 0)break;endargc = argc + 1;endif (argc == 1)disp(`Usage Error 2`);help seqReconerror(`Usage Error`);endobs = {varargin{1:argc-1}};%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% Read the knowledge baseif (isKnowledge(varargin{argc}) == 0)disp(`Usage Error 3');help seqReconerror(`Usage Error');endknl = varargin{argc};argc = argc + 1;sel = knl.sel;rep = knl.rep;if (length(rep) .sup..about. = length(obs))disp(`Usage Error 4`);help seqReconerror(`Usage Error`);endNumTrainRep = length(obs);for i = 1:NumTrainRepif (strcmp(knl.rep{i}, obs{i}.type) == 0)error(`Error: Corresponding observations and knowledge bases mustbe in the same representation`);endend%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% Process Representation and NumEigVectfor i = 1:2if (argc <= nargin)if (isSeqType(varargin{argc}))newrep = varargin{argc};argc = argc + 1;elseif (isstr(varargin{argc}) == 0)nEv = varargin{argc};argc = argc + 1;endendend%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% Process Remaining Optionswhile (argc <= nargin)if (isstr(varargin{argc}) == 0)disp(`Usage Error 6`);help seqReconerror (`Usage Error`);endswitch varargin{argc}case `betai`betai.sub.-- given = varargin{argc+1};argc = argc + 2;case `gausprior`lambdaGausPrior = varargin{argc+1};argc = argc + 2;case `pulse`pulse = 1;knlPulse = varargin{argc+1};seqIm = varargin{argc+2};pulse.sub.-- stacked.sub.-- rep = varargin{argc+3};pulse.sub.-- sensor.sub.-- pos = varargin{argc+4};pulse.sub.-- sensor.sub.-- ori = varargin{argc+5};pulse.sub.-- twosigmasq = varargin{argc+6};pulse.sub.-- cos.sub.-- power = varargin{argc+7};argc = argc + 8;case `core`reg{1} = varargin{argc+1};reg{2} = varargin{argc+2};reg{3} = 1;argc = argc + 3;case `clip`reg{1} = varargin{argc+1};reg{2} = varargin{argc+2};reg{4} = 1;argc = argc + 3;case `sum1`reg{5} = varargin{argc+1};reg{2} = varargin{argc+2};argc = argc + 3;case `reg`reg = {varargin{argc+1:argc+5}};argc = argc + 2;case `weights`weights = varargin{argc+1};argc = argc + 2;case `Mnewrep`Mnewrep = varargin{argc+1};argc = argc + 2;case `interactive`interactive = 1;corrLambda = varargin{argc+1};seqlxn = varargin{argc+2};argc = argc + 3;case `track`track = 1;seqEdge = varargin{argc+1};argc = argc + 2;case `ptcorr`ptcorr = varargin{argc+1};argc = argc + 2;case `trackcorr`trackcorr = varargin{argc+1};argc = argc + 2;case `energymin`energymin = 1;lamKE = varargin{argc+1};lamLenE = varargin{argc+2};argc = argc + 3;otherwisedisp(`Usage Error 5`);help seqReconerror(`Usage Error`);endndnums = sel.seq;ori = sel.ori;txtr = sel.texture;encParams = knl.encParams;%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% Encode the Training Matrix in the New Representation (if not%% supplied)if (isempty(Mnewrep))if ((strcmp(rep{1}, newrep)) & (isempty(Mnewrep)))Mnewrep = knl.enc{1};if (exists(knl, `mean`))MnewrepMean = knl.mean{1};endelsec = 1;for n = nums;for j = ori; for k = txtr;if (n > 10) nseq = seqsubSample(seqGet(n, newrep, j, k), 2);else nseq = seqGet(n, newrep, j, k); end if (isempty(Mnewrep)) Mnewrep = seqEncode(nseq, encParams); else Mnewrep = [Mnewrep seqEncode(nseq, encParams)]; end c = c + 1; endendend%% See if we're doing the mean alignmentif (exists(knl, `mean`))%% If so, subtract the mean of the M newrep . . .MnewrepMean = mean(Mnewrep, 2);Mnewrep = Mnewrep - MnewrepMean * ones(1, size(Mnewrep, 2));endendelse%% See if we're doing the mean alignmentif (exists(knl, `mean`))%% If so, subtract the mean of the M newrep . . .MnewrepMean = mean(Mnewrep, 2);Mnewrep = Mnewrep - MnewrepMean * ones(1, size(Mnewrep, 2));endend%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% Encode the observationsfor i = 1:NumTrainRepif (exists(obs{i}, `encoded`))Mobs{i} = obs{i}.encoded;elseMobs{i} = seqEncode(obs{i}, knl.encParams);endif (exists(knl, `mean`))Mobs{i} = Mobs{i} - knl.mean{i} * ones(1, size(Mobs{i}, 2));endend%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% Compute M'y, given the encoded observationsMty = 0;numTrainSnip = size(knl.enc{1}, 2);for i = 1:NumTrainRepMty = Mty + knl.weights(i) * knl.enc{i}' * Mobs{i};endif (nEv > min(size(knl.S)))nEv = min(size(knl.S));disp(sprintf(`seqRecon: Warning: Can only use %d eigenvectors`,nEv));end% Si is pinv(S) (S is diag, so it 1/diag(S))Si = spdiags(1./diag(knl.S), 0, size(knl.V,1), nEv);% Coeffs are coefficients of the eigen vectors.coeffs = knl.U(:,1:nEv)' * Mty;% betai are coeffiecients of training databetai = knl.V * (Si * coeffs);numSnip = size(betai, 2);if (exists(knl, `MtM`) == 0)disp(`seqRecon: Warning: MtM does not exist. Computing. (slow!)`);knl.MtM = knl.U * knl.S * knl.V';end%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% Minimization LoopMtM = knl.MtM;MtM0 = MtM;Mty0 = Mty;itering = 1;global seqINFOnmk = seqINFO.nummkr(obs{1}.subtype);%% Initializeinit = cell(numSnip,1);for i=1:numSnip; init{i}=cell(0); endlast.sub.-- t = 1;% MtM0, Mty0 contain the observation constraint (no correpondence% constraints)first.sub.-- iter = 1;while (itering)itering = 0;%% Initialize for this iterationMtM = MtM0;Mty = Mty0;M.sub.-- s = init;y.sub.-- s = init;weight.sub.-- s = cell(numSnip, 1);%% Include the point correspondence constraintsLMtM, Mty, M.sub.-- s, y.sub.-- s, weight.sub.-- s] = compConstr(MtM,Mty, 1, ptcorr,corrLambda, obs{1}, Mnewrep, nmk, encParams, numSnip, M.sub.-- s,y.sub.-- s,weight.sub.-- s);if (track)%% Track based on the initial constraintstrackcorr = seqTrack(seqEdge, trackcorr);%% Include the point correspondence constraints[MtM, Mty, M.sub.-- s, y.sub.-- s, weight.sub.-- s] = compConstr(MtM,Mty, 1,trackcorr, corrLambda, obs{1}, Mnewrep, nmk, encParams, numSnip,M.sub.-- s, y.sub.-- s, weight.sub.-- s);endif ((first.sub.-- iter == 1) & (isempty(betai.sub.-- given)==0))first.sub.-- iter = 0;betai = betai.sub.-- given;else%% Regularizingbetai = regularize(Mty, MtM, betai, nEv, reg{1:5});end%% ProjectingMrecon = Mnewrep * betai;if (exists(knl, `mean`))Mrecon = Mrecon + MnewrepMean * ones(1, size(Mrecon, 2));end%% Decodeseq = seqDecode(Mrecon, newrep, encParams, obs{1});%% Minimize with pulseif (pulse)%% Reset all the data matricesMtM = MtM0;Mty = Mty0;M.sub.-- s = init;y.sub.-- s = init;weight.sub.-- s = cell(numSnip, 1);s = knlPulse.encParams.fps;k = knlPulse.encParams.step;seqPulse = seqConvert(seq, `pulse`);encPulse = seqEncode(seqPulse, knlPulse.encParams);pulse.sub.-- u = knlpulse.U;%% Should be made (another) parameterobsLambda = 0;lamPulse = 100000;Sip = spdiags(1./diag(knlPulse.S), 0, size(knlPulse.V,1), nEv);VpSpi = knlPulse.V * Sip;for snip = 1:numSnippulse.sub.-- goal = seqConvert(seqCrop(seqIm, k*(snip-1)+1,k*(snip-1)+s), . . .`formc`, pulse.sub.-- sensor.sub.-- pos, pulse.sub.-- sensor.sub.-- ori,. . .pulse.sub.-- twosigmasq, pulse.sub.-- cos.sub.-- power);% bet0 = knlPulse.U(:,i:nEv)' * (knlpulse.enc{1}' * encPulse);bet0 = knlPulse.S(1:nEv,:) * knlPulse.V' * betai(:,snip);%% Add the correspondence constraints to the observation[MtM, Mty, M.sub.-- s, y.sub.-- s, weight.sub.-- s] = compConstr(MtM,Mty, obsLambda,. . .ptcorr, corrLambda, obs{1}, Mnewrep, nmk, encParams, numSnip, . . .M.sub.-- s, y.sub.-- s, weight.sub.-- s);%% Add the tracking constraints to the observation[MtM, Mty, M.sub.-- s, y.sub.-- s, weight.sub.-- s] = compConstr(MtM,Mty, 1,trackcorr, . . .corrLambda, obs{1}, Mnewrep, nmk, encParams, numSnip, . . .M.sub.-- s, y.sub.-- s, weight.sub.-- s);Mobs.sub.-- snip = Mobs;for i = 1:length(Mobs) Mobs.sub.-- snip{i} = Mobs{i}(:,snip);end%% Initialize the energy functionfEnergyInit(VpSpi, Mnewrep, . . .knl.enc, Mobs.sub.-- snip, knl.weights, . . .M.sub.-- s(snip), y.sub.-- s(snip), weight.sub.-- s(snip), . . .MtM(snip), Mty(:,snip), nEv, . . .knlPulse, pulse.sub.-- goal, pulse.sub.-- stacked.sub.-- rep, . . .pulse.sub.-- sensor.sub.-- pos, pulse.sub.-- sensor.sub.-- ori, . . .pulse.sub.-- twosigmasq, pulse.sub.-- cos.sub.-- power, . . .lamKE, lamLenE, lamPulse);options(1) = 1;options(2) = 0.001;options(3) = 0.001;options(14) = 100;bet = fminu(`fEnergy`, bet0, options, `fdEnergy`);betai(:,snip) = VpSpi * bet;end%% ProjectingMrecon = Mnewrep * betai;%% Decodeseq = seqDecode(Mrecon, newrep, encParams, obs{1});end%% Interactive Editingif (interactive)[ptcorr, trackcorr, itering, last.sub.-- t] = addConstraints(seq, seqIm,. . .ptcorr, trackcorr,last.sub.-- t);endend%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% Set fields of the result sequenceorigRep = obs{1}.type;seq.angle = obs{1}.angle;seq.subtype = obs{1}.subtype;seq.recfrom = origRep;seq.trainedOn = knl.sel;seq.trainType = knl.rep;seq.name = `Reconstruction of "` obs{1}.name `"`];file: /homes/leventon/Matlab/seqReconISSD.m% nEv = 50;pulse.sub.-- stacked.sub.-- rep = `stackedpulse.sub.-- no.sub.-- normals`;pulse.sub.-- sensor.sub.-- pos = sensor.sub.-- pos;pulse.sub.-- sensor.sub.-- ori = sensor.sub.-- ori;pulse.sub.-- twosigmasq = twosigmasq;pulse.sub.-- cos.sub.-- power = cos.sub.-- power;numSnip = 18;init = cell(numSnip,1);for i=1:numSnip; init{i}=cell(0); endS3i = spdiags(1./diag(knl3d.S), 0, size(knl3d.V,1), nEv);V3S3i = knl3d.V * 53i;Mnewrep = knl3d.enc{1};global M3dMeanM3dMean = knl3d.mean{1};nmk = 37;encParams = knl3d.encParams;s = knl3d.encParams.fps;k = knl3d.encParams.step;last.sub.-- t = 1;track = 0;% seq = seqCrop(seqGet(10), 1, seqSm.nframes);% seq = seqCrop(seqGet(10), 1, 113);% seq.scale = seqSm.scale;% enc = seqEncode(seq, `force`, 1);% betai = knl3d.V * S31(:,i:nEv) * knl3d.U(:,i:nEv)' * enc;%ptcorr = ptcorr.sub.-- bill;%trackcorr = trackcorr.sub.-- bill;ptcorr = ptcorr.sub.-- barish;trackcorr = { };interactive = 0;% betai0 = betai;% alpha3 = knl3d.S(1:nEv,:) * knl3d.V' * betai;itering = 1;while (itering)%% Reset all the data matricesMtM = zeros(sizeknl3d.S, 2));Mty = zeros(size(knl3d.S, 2), numSnip);M.sub.-- s = init;y.sub.-- s = init;weight.sub.-- s = cell(numSnip, 1);%% Interactive Editingif (interactive == 1)[ptcorr, trackcorr, itering, last.sub.-- t] = addConstraints(seq, seqIm,. . .ptcorr, trackcorr, last.sub.-- t);if (itering == 0), break; endelseitering = 0;end%% Add the correspondence constraints to the observation[MtM, Mty, M.sub.-- s, y.sub.-- s, weight.sub.-- s] = compConstr(MtM,Mty, 0, . . ./ - ptcorr, 1, seqSm, Mnewrep, nmk, encParams, numSnip, . ..M.sub.-- s, y.sub.-- s, weight.sub.-- s, M3dMean);if (track)%% Track based on the initial constraints[trackcorr, Models, Range, HausFrac, Certainty] = . . .seqTrack(seqEd, trackcorr, 0.25);addConstraints(seq, seqIm, ptcorr, trackcorr, last.sub.-- t);%% Add the tracking constraints to the observation[MtM, Mty, M.sub.-- s, y.sub.-- s, weight.sub.-- s] = compConstr(MtM,Mty, 1, trackcorr,. . .1, seqSm, Mnewrp, nmk, encParams, numSnip, . . .M.sub.-- s, y.sub.-- s, weight.sub.-- s, M3dMean);endfor snip = 1:2seqSmCrop = seqCrop(seqSm, k*(snip-1)+1, k*(snip-1)+s);% pulse.sub.-- goal = seqConvert(seqSmCrop, . . .% `formc`, pulse.sub.-- sensor.sub.-- pos, pulse.sub.-- sensor.sub.--ori, . . .% pulse.sub.-- twosigmasq, pulse.sub.-- cos.sub.-- power);fEnergyInit2(`PulseE`, 0, {knl3d, pulse.sub.-- goal, nEV,pulse.sub.-- stacked.sub.-- rep, . . .pulse.sub.-- sensor.sub.-- pos, pulse.sub.-- sensor.sub.-- ori, . . .pulse.sub.-- twosigmasq, pulse.sub.-- cos.sub.-- power}, . . .`PriorE`, 0.1, {S31}, . . .`ISSDE`, 0.05, {Mnewrep, V3S3i, seqSmCrop, 1, . . .knl3d.mean{1}, encParams, 1}, . . .`CorrE`, 0.5, {M.sub.-- s(snip), y.sub.-- s(snip), weight.sub.-- s(snip),. . .MtM(snip), Mty(:,snip), nEv, V3S3i});options(1) = 1;options(2) = 0.001;options(3) = 0.001;options(14) = 1000;% options(16) = 0.00001;alpha3(:,snip) = fminu(`fEnergy2`, alpha3(:,snip), options);betai(:,snip) = V3S3i * alpha3(:,snip);end%% ProjectingMrecon = Mnewrep * betai(:, i:snip);if (exists(knl3d, `mean`))Mrecon = Mrecon + knl3d.mean{1} * ones(1, size(Mrecon, 2));end%% Decodeseq = seqDecode(Mrecon, `3d`, encParams, seqSm);end% addConstraints(seq, seqIm, ptcorr, trackcorr, last.sub.-- t);ptcorrLast = ptcorr;trackcorrLast = trackcorr;returnnEv = 50;S31 = spdiags(1./diag(knl3d.S), 0, size(knl3d.V,1), nEv);V3S3i = knl3d.V * S3i;alpha3 = knl3d.S(1:nEv,:) * knl3d.V' * betai;betai(:,snip) = V3S31 * alpha3(:,snip);Mrecon = Mnewrep * betai(:, 1:snip);if (exists(knl3d, `mean`))Mrecon = Mrecon + knl3d.mean{1} * ones(1, size(Mrecon, 2));endseq = seqDecode(Mrecon, `3d`, encParams, seqSm);% nev i 1P 1C 1I fIter%--------------------------------% 50 1 0.1 0.5 0 500% 25 0 0.1 0.5 0 1000% 25 0 0.1 0.5 0.05 1000% 50 0 0.1 0.5 0.05 1000file: /homes/leventon/Matlab/seqReconPulse.mnEv = 50;pulse.sub.-- stacked.sub.-- rep = `stackedpulse.sub.-- no.sub.-- normals`;pulse.sub.-- sensor.sub.-- pos = sensor.sub.-- pos;pulse.sub.-- sensor.sub.-- ori = sensor.sub.-- ori;pulse.sub.-- twosigmasq = twosigmasq;pulse.sub.-- cos.sub.-- power = cos.sub.-- power;numSnip = 18;init = cell(numSnip,1);for i=1:numSnip; init{i}=cell(0); endS3i = spdiags(1./diag(knl3d.S), 0, size(knl3d.V,1), nEv);V3S3i = knl3d.V * S3i;Mnewrep = knl3d.enc{1};global M3dMeanM3dMean = knl3d.mean{1};nmk =37;encParams = kn13d.encParams;s = knl3d.encParams.fps;k = knl3d.encparams.step;last.sub.-- t = 1;track = 1;% seq = seqcrop(seqGet(10), i, seqsm.nframes);% seq = seqcrop(seqGet(10), 1, 113);% seq.scale = seqSm.scale;% enc = seqEncode(seq, `force`, 1);% betai = knl3d.V * S3i(:,1:nEv) * knl3d.U(:,i:nEv)' * enc;ptcorr = ptcorr.sub.-- bill;trackcorr = trackcorr.sub.-- bill;interactive = 0;% betai0 = betai;% alpha3 = kn13d.S(1:nEv,:) * knl3d.V' * betai;itering = 1;while (itering)%% Reset all the data matricesMtM = zeros(size(knl3d.S, 2));Mty = zeros(size(kn13d.S, 2), numSnip);M.sub.-- s = init;y.sub.-- s = init;weight.sub.-- s = cell(numSnip, 1);%% Interactive Editingif (interactive == 1)[ptcorr, trackcorr, itering, last.sub.-- t] = addconstraints(seq, seqIm,. . .ptcorr, trackcorr, last.sub.-- t);if (itering == 0), break; endelseitering = 0;end%% Add the correspondence constraints to the observation[MtM, Mty, M.sub.-- s, y.sub.-- s, weight.sub.-- s] = compConstr(MtM,Mty, 0, . . .ptcorr, 1, seqSm, Mnewrep, nmk, encParams, numSnip, . . .M.sub.-- s, y.sub.-- s, weight.sub.-- s);if (track)%% Track based on the initial constraints[trackcorr, Models, Range, HausFrac, Certainty] = . . .seqTrack(seqEd, trackcorr, 0.25);addConstraints(seq, seqIm, ptcorr, trackcorr, last.sub.-- t);%% Add the tracking constraints to the observation[MtM, Mty, M.sub.-- s, y.sub.-- s, weight.sub.-- s] = compConstr(MtM,Mty, 1, trackcorr,. . .1, seqSm, Mnewrep, nmk, encParams, numSnip, . . .M.sub.-- s, y.sub.-- s, weight.sub.-- s);endfor snip = 1:3pulse.sub.-- goal = seqConvert(seqCrop(seqSm, k*(snip-1)+1,k*(snip-1)+s),. . .`formc`, pulse.sub.-- sensor.sub.-- pos, pulse.sub.-- sensor.sub.-- ori,pulse.sub.-- twosigmasq, pulse.sub.-- cos.sub.-- power);fEnergyInit2(`PulseE`, 1, {knl3d, pulse.sub.-- goal, nEv,pulse.sub.-- stacked.sub.-- rep, . . .pulse.sub.-- sensor.sub.-- pos, pulse.sub.-- sensor.sub.-- ori, . . .pulse.sub.-- twosigmasq, pulse cos.sub.-- power}, . . .`PriorE`,1, {S3i}, . . .`CorrE`, 5, {M.sub.-- s(snip), y.sub.-- s(snip), weight.sub.-- s(snip), .. .MtM(snip), Mty(:,snip), nEv, V3S3i});options(1) = 1;options(2) = 0.001;options(3) = 0.001;options(14) = 500;alpha3(:,snip) = fminu(`Energy2`, alpha3(:,snip), options);betai(:,snip) = V3S3i * alpha3(:,snip);end%% ProjectingMrecon = Mnewrep * betai(:, 1:snip);if (exists(knl3d, `mean`))Mrecon = Mrecon + MnewrepMean * ones(1, size(Mrecon, 2));end%% Decodeseq = seqDecode(Mrecon, `3d`, encParams, seqSm);end% addConstraints(seq, seqIm, ptcorr, trackcorr last.sub.-- t);ptcorrLast = ptcorr;trackcorrLast = trackcorr;returnfile: /homes/leventon/Matlab/seqReconTry.m% function seq = seqReconTry(seqIm, knl3d)%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% DEFAULTS% Reconstruct in 3dnewrep = `3d`;% Using 50 EigenvectorsnEv = 50;% Assume no regularization {lambdas, nIter, core?, clip?, sum1}reg = {0 0 0 0 0};reg = {1e3 1 0 1 1e5};% Don't Tracktrack = 0;% Assume we have to encode the training data in the new representationMnewrep = [ ];% Assume we don't have any point correspondencesptcorr = cell(0);% Assume we don't have any tracking correspondencestrackcorr = cell(0);% Assume we don't want to energy minimizeenergymin = 0;% Default lambda for correlationscorrLambda = 1e6;% Assume we don't want interactiveinteractive = 1;% Assume we aren't using pulse minimizationpulse = 0;% Assume no gaussian prior on the coefficientslambdaGausPrior = 0;%lamLenE = 0;lamKE = 0;knl = knl3d;Mnewrep = knl.enc{1};MnewrepMean = knl.mean{1};%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% Minimization Loopobs{1} = seqIm;obs{1}.scale = 0.25;numSnip = 22;encParams = knl3d.encParams;MtM = zeros(size(Mnewrep,2), size(Mnewrep,2));MtM0 = MtM;Mty0 = zeros(size(Mnewrep,2), numSnip);itering = 1;global seqINFOnmk = 37;%% Initializeinit = cell(numSnip,1);for i=1:numSnip; init{i}=cell(0); endlast.sub.-- t = 1;% MtM0, Mty0 contain the observation constraint (no correpondence% constraints)while (itering)itering = 0;%% Initialize for this iterationMtM = MtM0;Mty = Mty0;M.sub.-- s= init;y.sub.-- s = init;weight.sub.-- s = cell(numSnip, 1);%% Include the point correspondence constraints[MtM, Mty, M.sub.-- s, y.sub.-- s, weight.sub.-- s] = compConstr(MtM,Mty, 0, ptcorr,corrLambda, obs{1}, Mnewrep, nmk, encParams, numSnip, M.sub.-- s,y.sub.-- s,weight.sub.-- s);if (track)%% Track based on the initial constraintstrackcorr = seqTrack(seqEdge, trackcorr);Include the point correspondence constraints[MtM, Mty, M.sub.-- s, y.sub.-- s, weight.sub.-- s] = compConstr(MtM,Mty, 1,trackcorr, corrLambda, obs{1}, Mnewrep, nmk, encParams, numSnip,M.sub.-- s, y.sub.-- s, weight.sub.-- s);end%% Regularizingif (xxx)ticbetai = regularize(Mty, MtM, betai, nEv, reg{1:5});tocend%% ProjectingMrecon = Mnewrep * betai;if (exists(kni, `mean`))Mrecon = Mrecon + MnewrepMean * ones(1, size(Mrecon, 2));end%% Decodeseq = seqDecode(Mrecon, newrep, encParams, obs{1});%% Minimize with pulseif (pulse)%% Reset all the data matricesMtM = MtM0;Mty = Mty0;M.sub.-- s = init;y.sub.-- s = init;weight.sub.-- s = cell(numSnip, 1);s = knlpulse.encParams.fps;k = knlpulse.encParams.step;seqPulse = seqConvert(seq, `pulse`);encPulse = seqEncode(seqPulse, knlPulse. encParams);pulse.sub.-- u = knlpulse.U;%% Should be made (another) parameterobsLambda = 0;lamPulse = 100000;Sip = spdiags(1./diag(knlPulse.S), 0, size(knlPulse.V,1), nEv);VpSpi = knlpulse.V * Sip;for snip = 1:numSnippulse.sub.-- goal = seqConvert(seqCrop(seqIm, k*(snip-1)+1,k*(snip-1)+s), . . .`formc`, pulse sensor.sub.-- pos, pulse.sub.-- sensor.sub.-- ori, . . .pulse.sub.-- twosigmasq, pulse.sub.-- cos.sub.-- power);% bet0 = knlPulse.U(:,1:nEv)' * (knlPulse.enc{1}' * encPulse);bet0 = knlpulse.S(1:nEv,:) * knlPulse.V' * betai(:,snip);%% Add the correspondence constraints to the observation[MtM, Mty, M.sub.-- s, y.sub.-- s, weight.sub.-- s] = compConstr(MtM,Mty, obsLambda,. . .ptcorr, corrLambda, obs{1}, Mnewrep, nmk, encParams, numSnip,. . .M.sub.-- s, y.sub.-- s, weight.sub.-- s);%% Add the tracking constraints to the observation[MtM, Mty, M.sub.-- s, y.sub.-- s, weight.sub.-- s] = compConstr(MtM,Mty, 1,trackcorr, . . .corrLambda, obs{1}, Mnewrep, nmk, encParams, numSnip, . . .M.sub.-- s, y.sub.-- s, weight.sub.-- s);Mobs.sub.-- snip = Mobs;for i = 1:length(Mobs)Mobs.sub.-- snip{i} = Mobs{i}(:,snip);end%% Initialize the energy functionfEnergyInit(VpSpi, Mnewrep, . . .knl.enc, Mobs.sub.-- snip, knl.weights, . . .M.sub.-- s(snip), y.sub.-- s(snip), weight.sub.-- s(snip), . . .MtM(snip), Mty(:,snip), nEv,knlPulse, pulse.sub.-- goal, pulse.sub.-- stacked.sub.-- rep, . . .pulse.sub.-- sensor.sub.-- pos, pulse.sub.-- sensor.sub.-- ori, . . .pulse.sub.-- twosigmasq, pulse.sub.-- cos.sub.-- power, . . .lamKE, lamLenE, lamPulse);options(1) = 1;options(2) = 0.001;options(3) = 0.001;options(14) = 100;bet = fminu(`fEnergy`, bet0, options, `fdEnergy`);betai(:,snip) = VpSpi * bet;end%% ProjectingMrecon = Mnewrep * betai;%% Decodeseq = seqDecode(Mrecon, newrep, encParams, obs{1});end%% Interactive Editingif (interactive)[ptcorr, trackcorr, itering, last.sub.-- t] = addConstraints(seq, seqIm,. . .ptcorr, trackcorr,last.sub.-- t);endend%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% Set fields of the result sequenceorigRep = obs{1}.type;seq.angle = obs{1}.angle;seq.subtype = obs{1}.subtype;seq.recfrom = origRep;seq.trainedOn = knl.sel;seq.trainType = knl.rep;seq.name = [`Reconstruction of "` obs{1}.name `"`];file: /homes/leventon/Matlab/seqRecon.sub.-- old.mfunction [seq, M, coeffs, betai, sic] = seqRecon(obs, knl, newrep,nEv, core, Mnewrep, forcei)if (nargin < 3)newrep = `3d`;endif (nargin < 4)nEv = 50;endif (nargin < 6)Mnewrep = [ ];endif (nargin < 5)core = [ ];endif (nargin < 6)force1 = 0;endsel = knl.sel;nums = sel.seq;ori = sel.ori;txtr = sel.texture;rep = knl.rep;encParams = knl.encParams;if (isempty(Mnewrep))c = 1;for n = nums;for j = ori;for k = txtr; nseq = seqGet(n, newrep, j, k); if (isempty(Mnewrep)) Mnewrep = seqEncode(nseq, encParams); else Mnewrep = [Mnewrep seqEncode(nseq, encParams)]; end c = c + 1;endendendendorigRep = obs.type;obs = seqconvert(obs, rep);Mobs = seqEncode(obs, encParams);S = knl.S;U = knl.U;V = knl.V;if (nEv > min(size(S)))nEv = min(size(S));endcoeffs = U(:,1:nEv)' * Mobs;Si = spdiags(1./diag(S), 0, size(V,1), nEv);if (0)SIZE.sub.-- Mnewrep = size(Mnewrep)SIZE.sub.-- V = size(V)SIZE.sub.-- Si = size(Si)SIZE.sub.-- coeefs = size(coeffs)endsic = Si * coeffs;betai = V * sic;if (isempty(core) == 0)if (force1)betai = corelCoeff(Mobs, knl, betai, nEv, core{1}, core{2}, 1e5);% betai = clip1Coeff(Mobs, knl, betai, nEv, core{1}, core{2}, 1e5);elsebetai = coreCoeff(Mobs, knl, betai, nEv, core{1}, core{2});endendM = Mnewrep * V * Si;% Mrecon = M * coeffs;Mrecon = Mnewrep * betai;if (strcmp(rep, `image`) .vertline. strcmp(rep, `motion`))offsets = obs.head.data2d;elseoffsets = [ ];endseq = seqDecode(Mrecon, newrep, encParams, offsets);seq.angle = obs.angle;seq.subtype = obs.subtype;seq.recfrom = origRep;seq.trainedOn = knl.sel;seq.trainType = knl.rep;seq.name = [`Reconstruction of "`obs.name `"`];file: /homes/leventon/Matlab/seqRender.mfunction seqr = seqRender(seq, angle, step, dnsamp)% step == 0: `Do all three`% step == 1: `Fit and Render (don't load images)`% step == 2: `Load images (don't fit or render)`% step == 3: `Fit only`global seqINFOif (isempty(seqINFO.defaults.display))error(`Need to set the display for Vtk. Use seqsetvtkDisplay`);endif (nargin < 2)angle = 0;endif (nargin < 3)step = 0;endif (nargin < 4)dnsamp = 2;endif (iscell(seq))if (strcmp(seq{1}, `Rendering Process`))angle = seq{3};seq = seq{2};step = 2;endend%% . . ./Vtk/renderguy.tcl Temp/person.mdl Temp/tmp.seqTemp/tmp-%03d.ppm90 1 1angle = round(angle * 180 / pi);rc = seqINFO.defaults.rendercmd;m = seqINFO.defaults.model;s = seqINFO.defaults.seq;f = seqINFO.defaults.framespec;model = seqINFO.repAngle.model;display ` seqINFO.defaults.display];if (step .sup..about. = 2)disp(`seqRender: Fitting Cylinders`);seq = seqConvert(seq, `ang`);saveang(model, seq.data, s);if (step == 3)returnendif (step == 0)cmd = sprintf(`%s %s %s %s %s %d 1 1`, rc, display, m, s, f,angle);elsecmd = sprintf(`%s %s %s %s %s %d 1 1 &`, rc, display, m, s, f,angle);enddisp(`seqRender: Rendering`);unix(cmd);seqr = {`Rendering Process`, seq, angle};endN = seq.nframes;if (step.sup..about. = 1)disp(`seqRender: Loading Images`);for 1=1:Ns = sprintf(f, i);data{i} = downsample(ppInRead(s), dnsamp);endseqr.type = `image`;seqr.name = [`Rendered ` seq.name];seqr.subtype = seq.subtype;seqr.data = data;seqr.angle = angle;seqr.texture = 0;seqr.nframes = N;% seqr.headpos = worldToCa m(headpos(seq));%% Still broken . . . Should rotate the head sequence . . .seqr.head = seq.head;endfile: /homes/leventon/Matlab/seqRotate.mfunction seq = seqRotate(seq, theta)if (theta == 0)returnendswitch seq.typecase `3d`A = seq.data;n = size(A, 1);A1 = A(2:n, :);B1 = [(cos(theta)*A1(:,1) + -sin(theta)*A1(:,2))(sin(theta)*A1(:,1) + cos(theta)*A1(:,2)) A1(:,3)];B = [A(1, :); B1]seq.data = B;seq.angle = seq.angle + theta;seq.angle = seq.angle +sign(seq.angle)*(2*pi)*floor(seq.angle/(2*pi));otherwiseerror(sprintf(`Cannot rotate Sequences of type %s.`, seq.type));endfile: /homes/leventon/Matlab/seqSelect.mfunction sel = seqselect(seqnums, ori, texture)% seqSelect Select out sequences to train on%% sel = seqSelect(which.sub.-- seq, which.sub.-- ori) selects out allsequences in% the% vector which.sub.-- seq in all orientations in the vector which.sub.--ori.%% For exmaple seqSelect(1:10, [0, pi/2]) selects all 10 sequences at% both% 0 and 90 degrees.global seqINFOif (nargin < 3)texture = seqINFO.defaults.texture;endif (nargin < 2)ori = seqINFO.defaults.orientation;endsel.seq = seqnums;sel.ori = ori;sel.texture = texture;file: /homes/leventon/Matlab/seqSetvtkDisplay.mfunction seqSetvtkDisplay(display)% seqSetvtkDisplay Set the display to use for seqRender%% seqsetvtkDisplay(display) where display is of the form% "computer:0"%global seqINFOseqINFO.defaults.display = display;file: /homes/leventon/Matlab/seqSmoothE.mfunction [e, ea, a] = seqsmoothE(seq, a)if (nargin < 2)a = mkr2ang(seq.data);ende = 0;ea = zeros(1, size(a,2)-4);for i = 8:2:26t = a(i,:);p = a(i+1,:);f = (cos(t).*cos(p)-1). 2 + (cos(t).*cos(p)). 2 + sin(t). 2;[ei, eas] = errSmooth(f);ea = ea + eas;e = e + ei;endfile: /homes/leventon/Matlab/seqSubSample.mfunction seqn = seqSubsample(seq, factor)seqn = seq;N = seq.nframes;seqn.nframes = floor(N/factor);Nn = length(1:factor:N);seqn.name = [`Subsampled` seq.name];global seqINFOnmk = seqINFO.nummkr(seq.subtype);switch seq.typecase {`3d', `3dc`, `2d`}data = seqn.data(2:size(seqn.data,1), :);data = reshape(data, nmk, N, size(seq.data,2));data = data(:, i:factor:N, :);seqn.data = [seq.data(1,:); reshape(data, nmk*Nn,size(seq.data,2))];seqn.data(1,1) = Nn;seqn.head = seqsubSample(seq.head, factor);case {`image`, `motion`, `form`, `edge`}seqn.data = seq.data(1:factor:N);seqn.head = seqSubSample(seq.head, factor);case {`head`}seqn.data2d = seq.data2d(1:factor:N, :);if (exists(seq, `data3d`))seqn.data3d = seq.data3d(1:factor:N, :);endendfile: /homes/leventon/Matlab/seqTorqueE.mfunction [e, ea, a] = seqTorqueE(seq, a)if (nargin < 2)a = mkr2ang(seq.data);ende = 0;ea = zeros(10, size(a,2)-1);c = 1;for i = 8:2:26t = a(i,:);p = a(i+1,:);f = (cos(t).*cos(p)-1). 2 + (cos(t).*cos(p)). 2 + sin(t). 2;for t = 1:length(f)-1ea(c, t) = ea(c, t) + (f(t+1)-f(t)) 2;endc = c+1;ende = mean(mean(ea));file: /homes/leventon/Matlab/seqTrack.mfunction [trackcor, Models, Range, HausFrac, Certainty] =seqTrack(seqEd, trackcor, scale)%%scale =1;Params.FracThresh = [0.85 0.7 0.6 0.5 0.4 0.4 0.4 0.4 0.4 0.4 0.40.4];Params.CertainThresh = [0.8 0.6 0.5 0.4 0.3 0.3 0.3 0.3 0.3 0.3 0.30.3];Params.ModelGenDil = ones(5);Params.ModelSizeFrac = 0.5;Params.SearchRange = 10;global seqINFOnmkr = seqINFO.nuiftmkr(seqEd. subtype);nFr = seqEd.nframes;Models = cell(nFr, nmkr);nCor = length(trackcor);ModelPoints = zeros(nFr, nmkr);FirstTime = zeros(nFr, nmkr);Range = cell(nFr, nmkr);Certainty = zeros(nFr, nmkr);HausFrac = zeros(nFr, nmkr);global G2G2 = trackcor;for i = 1:nCormkr = trackcor{i}.mkr;t = trackcor{i}.t;pos = round(seqEd.scale * trackcor{i}.pos / scale);model = imSelect(seqEd.data{t}, pos);npts = length(find(model==0));if (npts > 0)Models{t, mkr} = model;HausFrac(t, mkr) = trackcor{i}.hausfrac;FirstTime(t, mkr) = trackcor{i}.first;Range{t, mkr} = pos;Certainty(t, mkr) = trackcor{i}.cert;ModelPoints(t, mkr) = npts;elsedisp(sprintf(`Warning: Model for marker %d at time %d is empty!`,mkr, t));endendfor t = 2:nFrTRACKING.sub.-- AT.sub.-- TIME = tfor mkr = 1:nmkrif ((isempty(Models{t-1, mkr}) == 0) & (isempty(Range{t, mkr})) &(Certainty(t-1, mkr) > 0) & (isempty(Range(t-1,mkr))==0))p = Range{t-1, mkr}(i,:);del = Params.SearchRange;[pos, frac] = haus(Models{t-1, mkr}, seqEd.data{t}, `r`,p(1)-del, p(2)-del, p(1)+del, p(2)+del);fracHausFrac(t, mkr) = frac;Certainty(t, mkr) = Certainty(t-1, mkr) * frac;mp = min(ModelPoints(t-1, mkr), ModelPoints(FirstTime(t-1, mkr),mkr));if (frac < Params.FracThresh(cell(mp/50)))disp(`Lost!`) Certainty(t, mkr) = 0;endif (Certainty(t, mkr) < Params.CertainThresh(cell(mp/50)))disp(`Lost!`) Certainty(t, mkr) = 0;endFirstTime(t, mkr) = FirstTime(t-1, mkr);t0 = FirstTime(t, mkr);if (ModelPoints(t-1, mkr) > ModelPoints(t0, mkr)) Range{t, mkr} = [pos-1; (pos+size(Models{t-1, mkr}))] m = imSelect(seqEd.data{t}, Range{t, mkr}); Models{t, mkr} = autocrop(and(dilate(pad(Models{t-1,mkr}, 1,1, 1)==0, Params.ModelGenDil), m==0) == 0, 1);else Range{t, mkr} = [pos; (pos+size(Models{t-1, mkr})-1] m = imselect(seqEd.data{t}, Range{t, mkr}); Models{t, mkr} = autocrop(and(dilate(Models{t-1,mkr}==0,Params.ModelGenDil), m==0) == 0, 1);endModelPoints(t, mkr) = length(find(Models{t, mkr}==0));if (ModelPoints(t, mkr) / Modelpoints(t0, mkr) <Params.ModelSizeFrac)disp(`Lost2!`) Certainty(t, mkr) = 0;endendendif (t > 3000)superimp({Models{t, :}}, seqEd.data{t}, {Range{t, :}}, 1);pause(0.1);endendtrackcor = cell(0);c = 1;for t = 1:nFrfor mkr = 1:nmkrif (Certainty(t, mkr) > 0)trackcor{c}.mkr = mkr;trackcor{c}.t = t;trackcor{c}.pos = Range{t, mkr}/ seqEd.scale * scale;trackcor{c}.cert = Certainty(t, mkr);trackcor{c},hausfrac = HausFrac(t, mkr);trackcor{c}.first = FirstTime(t, mkr);mp = min(ModelPoints(t, mkr), ModelPoints(FirstTime(t, mkr),mkr));mincert = Params.CertainThresh(cell(mp/50));trackcor{c}.weight = (Certainty(t, mkr) - mincert)/(1-mincert);c = c + 1;endendendfile: /homes/leventon/Matlab/seqTrain.m% seqTrain Train on a set of sequences%% seqTrain(sel, rep, . . . options . . .)%, seqTrain(sel, reps, weights, . . . options . . .)% seqTrain(sel, knls, weights, . . .options. . .)% seq Train(sel, knls, weights, . . . options . . .)%% sel is a selection of sequences (see seqSelect)% rep -- is a representation to train on% (e.g. `3d`, `2d`, `ang`, `motion`, `image`, `form`,%`stick`, . . .)%% OPTIONS:% `fps`, fps "Frames Per Snippet" Default = 10% `step`, step "Step sizes between Snippets" Default = 5% `smooth`, sm "Amount of Smoothing" Default = 0% `downsample`, dnsamp "Amount of Downsampling" Default = 2% `headalign`, headalign "Whether or not to zero the head%position"% Default = 1 when rep == "image", "motion", "form"% Default = 0 otherwise% `submean`, submean "Whether or not to subtract the mean"% `keepenc`, keepenc "Whether to keep the encoded matrix" Def% = 1% `MtM`, MtM "Whether to do svd on M or on M'*M% 0 <==> svd(M);% Default: 1 <==> svd(M'*M);function knl = seqTrain(sel, repknl, varargin)if (nargin < 2)repknl = `3d`;endif (iscell(repknl))if (nargin ( 3)weights = ones(1,length(repknl));options = cell(0);elseweights = varargin{1};options = {varargin{2 :length(varargin)}};if (length(weights) .sup..about. = length(repknl))error(`Length of weights must be equal to length of rep/knl`);endendelseoptions = varargin;repknl = {repknl};weights = 1;endN = length(repknl);if (isstr(repknl{1}))rep = repknl;headalign = 0;for i = 1:Nswitch rep{i}case {`image`, `motion`, `form`}headalign = 1;endendnums = sel.seq;ori = sel.ori;txtr = sel.texture;encParams = args(options{1:length(options)});encParams = args(encParams, `fps`, 10, `step`, 5, `smooth`, 0,`downsample`, 2, `headalign`, headalign, `keepenc`, 1, `submean`, 0,`MtM`, 1);headalign = encParams.headalign;encParams.subtype = [ ];clear repknlfor i = 1:Nrepknl{i}.enc{1} = [ ];c = 1;for n = nums;for j = ori; for k = txtr;if (n > 10) seq = seqSubSample(seqGet(n, rep{i}, j, k), 2);else seq = seqGet(n, rep{i}, j, k); end if (isempty(encParams.subtype))encParams.subtype = seq.subtype; elseif (encParams.subtype .sup..about. = seq.subtype)error(`Cannot encode sequences of different subtypes!`);end if (isempty(repknl{i})) repknl{i}.enc{1} = seqEncode(seq, encParams); else repknl{i}.enc{i} = [repknl{i}.enc{i}seqEncode(seq,encParams)]; end c = c + 1; endendendif (encParams submean)repknl{i}.mean{1} = mean(repknl{i}.enc{i}, 2);repknl{i}.enc{1} = repknl{i}.enc{i} - . . .repknl{i}.mean{1} * ones(1, size(repknl{i}.enc{i},2));endendelsefor i = 1:Nrep{i} = repknl{i}.rep{1};endencParams = repknl{1}.encParams;endif ((encParams.MtM == 0) & (N .sup..about. = 1))error(`Cannot train on multiple representations if doing svd on M`);endif (encParams.MtM == 1)MtM = 0;knl.enc = cell(N);for i = 1:NMtM = MtM + weights(i) * repknl{i}.enc{1}' * repknl{i}.enc{1};if (encParams.keepenc)knl.enc{i} = repknl{i}.enc{1};endif (encParams.submean)knl.mean{i} = repknl{i}.mean{i};endend[U, S, V] = svd(MtM, 0);knl.MtM = MtM;elseM = 0;knl.enc = cell(1);M = weights(i) * repknl{i}.enc{1};if (encParams.keepenc)knl.enc{i} = repknl{i}.enc{1};endif (encParams.submean)knl.mean{i} = repknl{i}.mean{1};end% knl = M;% return[U, S, V] = svd(M, 0);endS = spdiags(diag(S), 0, size(U,1), size(V,1));knl.sel = sel;knl.rep = rep;knl.weights = weights;knl.U = U;knl.S = S;knl.V = V;knl.encParams = encParams;knl.headalign = encParams.headalign;file: /homes/leventon/Matlab/seqView.mfunction Iout = seqView(seq, frames, varargin)% seqView View a Sequence%% seqView(seq) views the sequence.% seqView(seq, frames) views selected frames of the sequence% (frames = [ ] --> all frames)% seqView(seq, frames, options)%% Options:% `viewang`, [az, el]% Set the view angle of the 3d display% [0 0] displays along optical axis (what the camera%sees)% [90 0] displays perpen to the optical axis.%% `titleval`, title% Place a title on the figure.%% `filespec`, filespec% Save each frame to a color postscript file.% "filespec" should be of the form "foo-%d.ps" or% "foo-%02d.ps", etc.%% `connect`, connect?% Whether or not to connect the 3d or 2d markers with%lines.% (connect? = 0 or 1)%% `headalign`, headalign?% Whether or not to align the head to the center of the%image.% (headalign? = 0 or 1)%% `leds`, ledlist% List of led numbers to highlight%% `axisval`, axis ([x0 x1 y0 y1 z0 z1] or [x0 x1 y0 y1])%% Set the axes of the display.% By default, the axes are set to tightly fit the data%% `snippetwise`, interval% If aligning the head, how often to re-align it% (every "interval" frames). For example, in our% training data, we align once per snippet, which% is every 10 frames. Setting interval to 10% will realign just as the training data does.%% `downsample`, level% Downsample the images by 2 level%% `smooth`, sigma% Smooth the images by a gaussian with sigma "sigma"%% `color`, color?% For the multi-orientation representations% (form and stick), whether to render multiple% images for each orientation or to render% one color image with different changes as different% orientations%% See Also: seqView2%% 10/10/97 wtf added viewing for pulse representation. little% orientation elernents, in addition to points.% to view a sequence with 2 pluses per segment, from theta = 0% over 20 frames:% seqpulse = seqConvert(seqCrop(seq2d,20), `pulse`, {0}, {2});% seqView(seqpulse, 1:20);global seqINFOif (nargin > 2)if (iscell (varargin{1}))args = varargin{1};elseargs = cons(varargin{:});endendif (nargin < 2)frames = [ ];endif (nargin < 3)args.array = 1;endswitch seq.typecase `3d`args = cons(args, `leds`, [ ], `viewang`, [ ], `axisval`, [ ],`titleval`, [ ], `filespec`, [ ], `connect`, 1, `stopani`, 0,`expand`, [0 0 0]);if (args.connect)connect = seqINFO.connections{seq.subtype};elseconnect = [ ];endif (args.stopani == 0)viewaoa(seq.data, connect, frames, args.leds, args.viewang, . . .args.axisval, args.titleval, args.filespec);elseif (isempty(frames)) colormap (gray (seq. nframes+6));else colormap(gray(length(frames)+6));endc = colormap;colormap(c(size(c,1)-3:-1:1, :));viewaoa.sub.-- hacked(seq.data, connect, frames, args.leds,args.viewang, . . .args.axisval, args.titleval, args.filespec,args.expand);endcase `2d`args = cons(args, `leds`, [ ], `axisval`, [ ], `titleval`, [ ],`filespec`, [ ], `connect`, 1);if (args.connect)connect = seqINFO.connections{seq.subtype};elseconnect = [ ];endviewaoa2(seq.data, connect, frames, args.leds, args.axisval,args.titleval, args.filespec);case `pulse`args = cons(args, `leds`, [ ], `axisval`, [ ], `titleval`, [ ],`filespec`, [ ], `connect`, 0);if (args.connect)connect seqINFO.connections{seq.subtype};elseconnect = [ ];end% prepare seq.data. Need to append the cells, one per time frame,% into one large array. And add header info.data.sub.-- arr = zeros(seq.nframes*size(seq.data{1},1)+1, 4);data.sub.-- arr(1,:) = [seq.nframes, size(seq.data{1},1), 0, 0];for t = 1:seq.nframesdata.sub.-- arr(1+(t-1)*size(seq.data{1},1) +(1:size(seq.data{1},1)),1:2) = seq.data{t,1};data.sub.-- arr(1+(t-1)*size(seq.data{1},1) +(1:size(seq.data{1},1)),3:4) = seq.data{t,2};endviewaoa2(data.sub.-- arr, connect, frames, args.leds, args .axisval,args.titleval, args.filespec);case {`image`, `edge`}args = cons(args, `titleval`, [ ], `filespec`, [ ], `headalign`, 0,`snippetwise`, 1, `downsample`, 0, `smooth`, 0);nfr = length(seq.data);if (isempty(frames))frames = 1:nfr;endhold off;colormap(gray);if (args.headalign)for t = framesif (mod(t-1, args.snippetwise) == 0) posx = seq.head.data2d(t,1); posy = seq.head.data2d(t,2);endI = imcrop(seq.data{t}, round(posy)-40, round(posx)-40, 80,80);if (size(1, 3) == 1) imagesc(smooth(downsample(I, args.downsample),args.smooth));else rgbimagesc(smooth(downsample(I, args.downsample),args.smooth));endaxis image;title(args.titleval);if (isempty(args.filespec) == 0) fn = sprintf(args.filespec, t); print(`-dpsc`, fn);endif (length(frames) > 1) pause(0.1);endendelsefor t = framesI = seq.data{t};if (size(I, 3) == 1) images(smooth(downsample(I, args.downsample),args.smooth));else rgbimagesc(smooth(downsample(I, args.downsample),args.smooth));endaxis image;title(args.titleval);if (isempty(args.filespec) == 0) fn = sprintf(args.filespec, t); print(`-dpsc`, fn);endif (length(frames) > 1) pause(0.1);endendendcase `formc`args = cons(args, `titleval`, [ ], `filespec`, [ ], `headalign`, 0,`snippetwise`, 1, `downsample`, 0, `smooth`, 0);nfr = length(seq.data);if (isempty(frames))frames = 1:nfr;endhold off;colormap (gray);forr t = fraxnesI = seq.data{t};imagesc(reshape(I, 10) 8));axis image;title(args.titleval);if (isempty(args.filespec) == 0) fn = sprintf(args.filespec, t); print(`-dpsc`, fn);endif (length(frames) > 1) pause(0.1);endendcase `motion`args = cons(args, `titleval`, [ ], `filespec`, [ ], `headalign`,0, `snippetwise`, 1);nfr = length(seq.data);if (isempty(frames))frames = 1:nfr;endhold off;colormap(`default`);if (args.headalign)args = cons(args, `titleval`, [ ], `filespec`, [ ], `headalign`,0, `snippetwise`, 1, `downsample`, 2, `smooth`, 1);for t = frames if (mod(t-1, args.snippetwise) == 0) posx = seq.head.data2d(t,1); posy = seq.head.data2d(t,2); end I = imCrop(seq.data{t}, round(posy)-40, round(posx)-40, 80,80); I22 = [I(:,:,1) I(:,:,2); I(:,:,3) I(:,:,4)]; imagesc(smooth(downsample(I22,args.downsample), args.smooth)); caxis([-pi pi]);axis image; hold on f = 2 args.downsample; plot([0 2*size(I, 1)]/f+0.5, [size(I,2) size(I,2)]/f+0.5,`k`);/ - plot([size(I,1) size(I,1)]/f+0.5, [0 2*size(I,2)]/f+0.5, `k`);hold offtitle(args.titleval);if (isempty(args.filespec) == 0) fn = sprintf(args.filespec, t); print(`-dpsc`, fn); end if (length(frames) > 1) pause(0.1); endendelseargs = cons(args, `titleval`, [ ], `filespec`, [ ], `headalign',0, `snippetwise`, 1, `downsample`, 0, `smooth`, 0);for t = frames I22 = [seq.data{t}(:,:,i) seq.data{t}(:,:,2); seq.data{t}(:,:,3) seq.data{t}(:,:,4)]; imagesc(smooth(downsample(I22,args.downsample), args.smooth)); caxis([-pi pi]);axis image; title(args.titleval); if (isempty(args.filespec) == 0) fn = sprintf(args.filespec, t); print(`-dpsc`, fn); end if (length(frames) > 1) pause(0.1); endendendcase `form`args = cons(args, `titleval`, [ ], `filespec`, [ ], `headalign`, 0,`snippetwise`, 1, `downsample`, 0, `smooth`, 0, `color`, 1);nfr = length(seq.data);if (isempty(frames))frames = 1:nfr;endhold off;colormap(gray);if (args.headalign)for t = frames if (mod(t-1, args.snippetwise) == 0) posx = seq.head.data2d(t,1); posy = seq.head.data2d(t,2); end I = imcrop(seq.data{t}, round(posy)-40, round(posx)-40, 80,80); I22 = ]I(:,:,1) I(:,:,2); I(:,:,3) I(:,:4)];if (args.coior) rgbimagesc(downsample(smooth(I ,args.smooth),args.downsample)); else imagesc(downsample(smooth(I22,args.smooth),args.downsample));endaxis image; hold off title(args.titleval); if (isempty(args.filespec) == 0) fn = sprintf(args.filespec, t); print(`-dpsc`, fn); end if (length(frames) > 1) pause(0.1); endendelsefor t = frames I = seq.data{t}; I22 = [I(:,:,1) I(:,:,2); I(:,:,3) I(:,:,4)];if (args.color) rgbimagesc(downsample(smooth(I,args.smooth),args.downsample)); else imagesc(downsample(smooth(I22,args.smooth),args.downsample));endaxis image; title(args.titleval); if (isempty(args.filespec) == 0) fn = sprintf(args.filespec, t); print (`-dpsc`, fn); end if (length(frames) > 1) pause(0.1); endendendcase `stick`args = cons(args, `titleval`, [ ], `filespec`, [ ], `headalign`, 0,`snippetwise`, 1, `downsample`, 0, `smooth`, 0, `color`, 1);nfr = length(seq.data);if (isempty(frames))frames = 1:nfr;endhold off;colormap(gray);if (args.headalign)for t = frames if (mod(t-1, args.snippetwise) == 0) posx = seq.head.data2d(t,1); posy = seq.head.data2d(t,2); end I = imCrop(seq.data{t}, round(posy)-40, round(posx)-40, 80,80); I22 = [I(:,:,1) I(:,:,2); I(:,:,3) I(:,:,4)];if (args.color) rgbimagesc(transpose2(downsample(smooth(I,args.smooth),args.downsample))); else imagesc(downsample(smooth(I22,args.smooth),args.downsample));endaxis image; hold off title(args.titleval); if (isempty(args.filespec) == 0) fn = sprintf(args.filespec, t); print(`-dpsc`, fn); end if (length(frames) > 1) pause(0.1); endendelsefor t = frames I = seq.data{t}; I22 = [I(:,:,1) I(:,:,2); I(:,:,3) I(:,:,4)];if (args.color) rgbimagesc(transpose2(downsample(smooth(I, args.smooth),args.downsample))); else imagesc(downsample(smooth(I22,args.smooth),args.downsample));endaxis image; title(args.titleval); if (isempty(args.filespec) == 0) fn = sprintf(args.filespec, t); print(`-dpsc`, fn); end if (length(frames) > 1) pause(0.1); endendendcase `ang`viewang(seq.model, seq.data, frames);otherwiseerror(`Bad Representation`);endif (nargout > 0)Iout = 1;endhold offfile: /homes/leventon/Matlab/seqview2.mfunction Iout = seqview2(seq, frames, args)global seqINFOif (nargin < 2)frames = [];endif (nargin < 3)args.array = 1;endif (iscell(seq) == 0)seq = {seq};endNseq = length(seq);reps.r3d = 0;reps.r2d = 0;reps.image = 0;reps.motion = 0;reps.angle = 0;maxT = 0;args = cons(args, `leds`, [ ], `viewang`, [ ], `axisval`, [ ],`titleval`, [ ], `filespec`, [ ], `connect`, 1, `headalign`, 0,`snippetwise`, 1, `subplots`, 0, `downsample`, 0, `smooth`, 0);scale = 1;for i = 1:NseqmaxT = max(maxT, seq{i}.nframes);conns{i} = [ ];nmk{i} = 0;switch seq{i}.typecase `3d`reps.r3d = 1;conns{i} = seqINFO.connections{seq{i}.subtype};nmk{i} = seq{i}.data(1,2);case `2d`reps.r2d = 1;conns{i} = seqINFO.connections{seq{i}.subtype};nmk{i} = seq{i}.data(1,2);case `motion`reps.motion = 1;case `image`reps.image = 1;if (exists(seq{i}, `scale`)) scale = seq{i}.scale;endcase `ang`reps.angle = 1;endif (args.connect == 0)conns{i} = [ ];endendif (args.subplots == 0)if (reps.angle == 1)if (sum([reps.motion reps.image reps.r2d]) .sup..about. = 0)args.subplots = 1;endendif (reps.image == 1)if (reps.motion == 1)args.subplots = 1;endendendif (reps.image == 1)disp = `image`;elseif (reps.motion == 1)disp = `motion`;elseif (reps.r2d == 1)disp = `2d`;elsedisp = `3d`;endif (args.subplots == 1)rnplots(1) = cell(sqrt(Nseq));rnplots(2) = cell(Nseq/rnplots(1));subplot(rnplots(1), rnplots(2), 1);endcolori = `brbgmckybrgmckybrgmckybrgmckybrgmcky`;if (isempty(frames))frames = 1:maxT;elsej = 1;for i=1:length(frames)if (frames(i) <= maxT)nframes(j) = frames(i);j = j + 1;endendframes = nframes;endfor t = framesif (args.subplots == 1)for i = i:Nseqsubplot(rnplots(1), rnplots(2), i);hold offendelsehold offendfor i = 1:Nseqif (args.subplots == 1)disp = seq{i}.type;subplot(rnplots(1), rnplots(2), i);endswitch seq{i}.typecase `3d` b = (t-1i)*nmk{i}+2;e = t*nmk{i}+1; switch dispcase `3d` p = seq{i}.data(b:e, :); plot3(p(:,1), p(:,2), p(:,3), [colori(i) `0`]);hold onfor j = args.leds p = seq{i}.data(j+b, :); plot3(p(1), p(2), p(3), [colori(i) `*`]); end for j = 1:size(conns{i}) p1 = seq{i}.data(b+conns{i}(j,1), :); p2 = seq{i}.data(b+conns{i}(j,2), :); drawseg3(p1, p2, colori(i)); endcase `2d` p = seq{i}.data(b:e, :); plot(p(:,1), p(:,3), [colori(i) `o`]);for j = args.leds p = seq{i}.data(j+b, :); plot3(p(1), p(3), [colori(i) `*`]); end for j = 1:size(conns{i}) p1 = seq{i}.data(b+conns{i}(j,1), :); p2 = seq{i}.data(b+conns{i}(j,2), :); plot([p1(1) p2(1)], [p1(3) p2(3)], colori(i)); endcase `image` if (exists(seq{i}, `scale`)) myscale = scale / seq{i}.scale; else myscale = scale; end p = seq{i}.data(b:e, :); p = worldToCam(p); plot(myscale*p(:,1), myscale*p(:,2), [colori(i) `o`]);hold onfor j = args.leds p = seq{i}.data(j+b, :); p = worldToCam(p); plot3(myscale*p(:,1), myscale*p(:,2), [colori(i) `*`]); end for j = 1:size(conns{i}) p1 = seq{i}.data(b+conns{i}(j,1), :); p2 = seq{i}.data(b+conns{i}(j,2), :); p1 = worldToCam(p1); p2 = worldToCam(p2); plot(myscale*[p1(1) p2(1)], myscale*[p1(2) p2(2)],colori(i)); endcase `motion`endcase `2d`b = (t-1)*nmk{i}+2;e = t*nmk{i}+1;switch disp case `2d` p = seq{i}.data(b:e, :); plot(p(1), p(2), [colori(i) `o`]);for j = args.leds p = seq{i}.data(j+b, :); plot3(p(1), p(2), [colori(i) `*`]); end for j = 1:size(conns{i}) p1 = seq{i}.data(b+conns{i}(j,i), :); p2 = seq{i}.data(b+conns{i}(j,2), :); plot([p1(i) p2(i)], [pi(2) p2(2)], colori(i)); endcase `image`case `motion`endcase `image`colormap(gray);if (args.headalign) if (mod(t-1, args.snippetwise) == 0) posx = seq{i}.head.dat.sub.-- d(t,1); posy = seq{i}.head.dat.sub.-- d(t,2); end I = imcrop(seq{i}.data{t}, round(posy)-40, round(posx)-40, 80, 80); else I = seq{i}.data{t};endif (size(I,3)==1) imagesc(downsample(I, args.downsample));else rgbimagesc(downsample(I, args .downsample));% imagesc(downsample(I(:,:,1), args.downsample));endaxis image;case `motion`% colormap(`default`);if ((args.headalign) & (exists(seq{i}, `headpos`))) if (mod(t-1, args.snippetwise) == 0) posx = seq{i}.head.dat.sub.-- d(t,1); posy = seq{i}.head.dat.sub.-- d(t,2); end I = imCrop(seq{i}.data{t}, round(posy)-40, round(posx)-40, 80, 80); I22 = [I(:,:,1) I(:,:,2); I(:,:,3) I(:,:4)]; imagesc(smooth(downsample(I22,args.downsample), args.smooth)); caxis([-pi pi]);axis image; hold on f = 2 args.downsample; plot([0 2*size(I, 1)]/f+.5, [size(I,2) size(I,2)]/f+.5, `k`); plot([size(I,1) size(I,1)]/f+.5, [0 2*size(I,2)]/f+.5, `k`); else I = [seq{i}.data{t}(:,:,1) seq{i}.data{t}(:,:,2); seq{i}.data{t}(:,:,3) seq{i}.data{t}(:,:,4)]; imagesc(downsample(I,args.downsample)); caxis([-pi pi]);axis image; endcase `ang`endtitle(args.titleval);hold onendxlabel(`x`); ylabel(`y`); zlabel(`z`);if (length(frames) > 1)pause(0.1);endif (isempty(args.filespec) == 0)fn = sprintf(args.filespec, i);print(`-dpsc`, fn);endendfile: /homes/leventon/Matlab/seqtest.m%% seqtest.m%% descend on solutions of body data. This is analogous%% to the toy problem test of test.m, except for body data.%%%% Sept. 29, 1997 freeman created.% oct. 1, 1997 wtf test plan: do things at just 1 time frame.% see if the descent works. Then increase to more time.% oct. 10, 1997 added seqView, and seqConvert modifications to% include pulse representation. Present problem: calc of form and% deriv response takes 8 seconds per frame, for 2 pulses per segment.% Just form only takes: 0.25 seconds per frame.% to do: run with a high dimensional search variable x% have grad function check that the f.sub.-- body evaluation one just% called was at the right position. else, call f.sub.-- body again.% remove extra zeros stored in the derivs for nori = 1.% put in form information.% oct. 18, 1997 wtf% %% formb contains both the form and the derivative information.% tic% seqR = seqConvert(seqpulse, `formb`);% toc% %% takes 160 seconds, with derivatives. 2 seconds without% derivatives.% %% omitting the exponential calculation from getsD, gives 160% seconds,% %% also.% % With one orientation, takes a lot less time.%% oct. 20, 1997 added stacked.sub.-- rep variable. Remember tmpsetup.m% Starting from a reasonable starting point (mean of s*v' values for% all eigenvectors), converges in 30 sec for 10-d vector.% oct. 20, 1997 yet to do: debug why the orientation version doesn't%% work. ie, doesn't do descent properly in the bet coordinates when% I operate on sensor orientation responses and pulse orientations,% too. ie, for stackedpulse.sub.-- no normals, it works, forstackedpulse,% it doesn't. Well, sometimes it does.% TODO: get rid of the "tmpsetup.m" file that I use to pass% variables into seqConvert.m% See bottom of file debug case for where the derivatives don't% work for the orientation case. Something's wrong somewhere . . .clear%% initializationsseqInitdisp(`Get sequence number 4`);seq3d = seqGet(4);seq2d = seqConvert(seq3d, `2d`);DISPLAY = 0;if (DISPLAY == 1)figure(1);subplot(2,1,1)seqView(seq3d, 1:1);subplot(2,1,2)seqView(seq2d, 1:1, `connect`, 0);% display this in 2-d.viewaoa2(seq2d.data);endif (DISPLAY == 1)figure(2)end% view from 0 degrees, 2 pulses per segment.%% seqpulse = seqConvert(seq2d, `pulse`, {0}, {2});seqpulse = seqConvert(seq2d, `pulse`, 2);if (DISPLAY == 1)seqView(seqpulse);end%% Where should these initializations go? wtf oct. 21 97%% set-up array of sensors%% sensor.sub.-- rows = 8;sensor.sub.-- cols = 10;%% evenly spaced sensors[tmp.sub.-- x, tmp.sub.-- y] = meshgrid(1:sensor.sub.-- rows,1:sensor.sub.-- cols) ;sensor.sub.-- pos = [reshape(tmp.sub.-- x, 1, sensor.sub.-- rows*sensor.sub.-- cols); . . .reshape(tmp.sub.-- y, 1, sensor.sub.-- rows*senso.sub.-- cols)];sensor.sub.-- pos = repmat([85;50],1,sensor.sub.-- rows*sensor.sub.--cols) + sensor.sub.-- pos;%% standard deviation of sensor gaussian response, in pixels.twosigmasq = 2 * 4 2;initSensors%% setup for only one orientation.num.sub.-- ori = 1;cos.sub.-- power = 0;%% evenly spaced between 0 and 180. For one ori, just fillers.tmp.sub.-- ori = (0:(num.sub.-- ori-1))*pi/num.sub.-- ori;sensor.sub.-- ori = [cos(tmp.sub.-- ori); sin(tmp.sub.-- ori)];stacked.sub.-- rep = `stackedpulse.sub.-- no.sub.-- normals`;%% stacked.sub.-- rep = `stackedpulse`;%% get observationssnip = 8;seq1 = seqCrop(seqpulse,1,8);seq.sub.-- goal = seqConvert(seq1, `formc`, sensor.sub.-- pos,sensor.sub.-- ori,twosigmasq, cos.sub.-- power);all = [ ];nseqmax = 10;for nseq = 1:nseqmaxseq = seqCrop(seqpulse,snip*(nseq-1)+1,snip*nseq);if (DISPLAY == 1)seqView(seq);endseq = seqConvert(seq,stacked.sub.-- rep, sensor.sub.-- pos, sensor.sub.--ori, twosigmasq,cos.sub.-- power);all = [all, seq.data];end[u,s,v] = svd(all,0);bet = s * v';%% you start at one body position, and arrive at the other, as you%% should.tic[bet.sub.-- stop,options] = . . .fminu(`f.sub.-- body.sub.-- bet`, mean(bet,2), [ ], `grad.sub.-- body.sub.-- bet`, u, . . . seq.sub.-- goal, stacked.sub.-- rep, sensor.sub.-- pos, sensor.sub.-- ori, twosigmasq,cos.sub.-- power);tocdisp([`num iterations: ` num2str(options(10))]);fprintf(1,` goal coefficients: %f.backslash.n`, bet(:,1))fprintf(1,` .backslash.n`)fprintf(1,` found coefficients: %f.backslash.n`, bet.sub.-- stop)%% Oct. 31, 1997%% this shows how to convert an image sequence to `formc`%% representation. Note: whether the data that comes out of this%% is reasonable or not, I haven't tested yet.disp(`Get the sequence as an image`);seqim = seqGet(9, `image`);disp(`And view it.`);seqView(seqIm)disp(`Convert this image to a form representation`);seq.sub.-- formc = seqConvert(seqIm, `formc`, sensor.sub.-- pos,sensor.sub.-- ori,twosigmasq, cos.sub.-- power);%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% debugging: test that the derivatives all work.%% (they do, oct. 16, 1997 wtf).% testing of the derivativesf.sub.-- body.sub.-- bet(:,2), u, seq.sub.-- goal, stacked.sub.--,sensor.sub.-- pos, sensot.sub.-- ori,twosigmasq, cos.sub.-- power);out = grad.sub.-- body.sub.-- bet(bet(:,2), u, seq.sub.-- goal,stacked.sub.-- rep, sensor.sub.-- pos,sensor.sub.-- ori, twosigmasq, cos.sub.-- power);del = zeros(size(bet):,2)));del(1,1) = 0.0001;%% yes, this all agrees, as it should(f.sub.-- body.sub.-- bet(bet(:,2)+del, u, seq.sub.-- goal,stacked.sub.-- rep, sensor.sub.-- pos,sensor.sub.-- ori, twosigmasq, cos.sub.-- power) - . . .f.sub.-- body.sub.-- bet(bet(:,2), u, seq.sub.-- goal, stacked.sub.--rep, sensor.sub.-- pso,sensor.sub.-- ori, twosigmasq, cos.sub.-- power)) / 0.0001% this also works. for no ori case.f.sub.-- body.sub.-- bet(bet(:,2), u, seq.sub.-- goal, stacked.sub.--rep, sensor.sub.-- pso, sensot.sub.-- ori,twosigmasq, cos.sub.-- power);out = grad.sub.-- body.sub.-- bet(bet(:,2), u, seq.sub.-- goal,stacked.sub.-- rep, sensor.sub.-- pos,sensor.sub.-- ori, twosigmasq, cos.sub.-- power);del = 0.0001 * rand(size(bet(:,2)));%% yes, this all agrees, as it shouldnum = (f.sub.-- body.sub.-- bet(bet(:,2)+del, u, seq.sub.-- goasl,stacked.sub.-- rep, sensor.sub.-- pos,sensor.sub.-- ori, twosigmasq, cos.sub.-- power) - . . .f.sub.-- body.sub.-- bet(bet(:,2), u, seq.sub.-- goal, stacked.sub.--rep, sensor.sub.-- pos,sensor.sub.-- ori, twosigmasq, cos.sub.-- power))anal = out' * del% test the derivative. These results are in perfect agreement% for the stackedpulse.sub.-- no.sub.-- normals case, but with.sup..about. 10-50% error% for the stackedpulse case.results = [ ];for j = 2:nseqmax% initializef.sub.-- body.sub.-- bet(bet(: ,j), u, seq.sub.-- goal, stacked.sub.--rep, sensor.sub.-- pos,sensor.sub.-- ori, twosigmaasq, cos.sub.-- power);% compute gradientout = grad.sub.-- body.sub.-- bet(bet(:,j), u, seq.sub.-- goal,stacked.sub.-- rep, sensor.sub.-- pos,sensor.sub.-- ori, twosigmasq, cos.sub.-- power);del = 0.0001 * rand(size(bet(:,j)));num = (f.sub.-- body.sub.-- bet(bet(:,j)+del, u, seq.sub.-- goal,stacked.sub.-- rep,sensor.sub.-- pos, sensor.sub.-- ori, twosigmasq, cos.sub.-- power) - . .. f.sub.-- body.sub.-- bet(bet(:,j), u, seq.sub.-- goal, stacked.sub.-- rep, sensor.sub.-- pos, sensor.sub.-- ori, twosigmasq, cos.sub.-- power)) ;anal = out' * del;results = [results; num, anal];endfile: /homes/leventon/Matlab/setModelData.mfunction model = setModelData(model, data)%% This is a mess. Hacked from mcylmodel.% data = [0; 0; 0; data];% data(4:6) = [0 0 0]';model{2} = data(1);model{3} = data(2);model{4} = data(3);model{5} = data(4);model{6} = data(5);model{7} = data(6);model{8}{1}{10} = data(7);model{8}{1}{11}{1}{8} = data(8);model{8}{1}{11}{1}{9} = data(9);model{8}{1}{11}{1}{11}{1}{8} = data(10);model{8}{1}{11}{1}{11}{1}{9} = data(11);model{8}{1}{11}{2}{8} = data(12);model{8}{1}{11}{2}{9} = data(13);model{8}{1}{11}{2}{11}{1}{8} = data(14);model{8}{1}{11}{2}{11}{1}{9} = data(15);model{8}{2}{11}{1}{8} = data(16);model{8}{2}{11}{1}{9} = data(17);model{8}{2}{11}{1}{11}{1}{8} = data(18);model{8}{2}{11}{1}{11}{1}{9} = data(19);model{8}{2}{11}{2}{8} = data(20);model{8}{2}{11}{2}{9} = data(21);model{8}{2}{11}{2}{11}{1}{8} = data(22);model{8}{2}{11}{2}{11}{1}{9} = data(23);model{8}{2}{11}{1}{11}{1}{11}{1}{8} = data(24);model{8}{2}{11}{1}{11}{1}{11}{1}{9} = data(25);model{8}{2}{11}{2}{11}{1}{11}{1}{8} = data(26);model{8}{2}{11}{2}{11}{1}{11}{1}{9} = data(27);model{8}{1}{11}{3}{10} = data(28);file: /homes/leventon/Matlab/showConstraints.mfunction showConstraints(seq, seqIm, ptcor, trackcor, frames)%% [ptcor, trackcor, itering, t] = addConstraints(seq, seqIm)% Starting with no constraints, the user interactively adds point or% tracking constraints. Seq is the reconstructed 3d sequence.% seqIm is the Image sequence. This function returns the% correspondences in two cells, "ptcor" for the point% correspondences,% and "trackcor" for the tracking correspondences.%% The return val itering tells if the user wants to continue% interating% through the svd-loop. (Whether the user clicked "Done" or "Cancel"%% and whether the user added anything from the last iteration).%% The return val t stores what time step the user left off at.%% [ptcor, trackcor, itering, t] = addConstraints(seq, seqIm, ptcor,% trackcor, t)% The ptcor and trackcor inputs allow for constraints to be% appended.% The t input determines at what time frame the interaction starts.% This function has pretty confusing control structure, but% matlab doesn't have the "continue" command . ..if (exists(seq, `scale`) == 0), seq.scale = 1; endif (exists(seqIm, `scale`) == 0), seqIm.scale = 1; endscale = seqIm.scale / seq.scale;if (nargin < 5) t = 1:min(seqIm.nframes,seq.nframes); endn = size(seq.data, 1);d = seq.data(2:n, :);nFr = min(seq.nframes, seqIm.nframes);d = reshape(d', 3, 37, nFr);d = transpose2(d);if (nargin < 4) trackcor = cell(0); endif (nargin < 3) ptcor = cell(0); end%% Dray Guy #1seqDiag = seqCrop(seqGet(6), 9, 9);guy1seq = seqconvert(seqDiag, `2d`);guy1dat = guy1seq.data(2:38,:);guy2seq = seqconvert(seqRotate(seqDiag, pi/2), `2d`);guy2dat = guy2seq.data(2:38,:);%% Start off by "showing" the already-established linksshowing = 1;%% We haven't made any changes yet . . .itering = 0;%% (There's no "do while" loop, so . . .)notdone = 1;%% There's no continue (only break), so . . .continue = 0;for t = frames%% Overlay the sequence on the image at time tseqView2({seqIm, seq}, t);axis off% Transfer the 3d data from world to camera coordinatesd2 = worldToCam(d(:,:,t));if (showing)%% Display all the Point Correspondencesfor i = 1:length(ptcor)if (ptcor{i}.t == t) mkr = ptcor{i}.mkr; p2 = ptcor{i}.pos([2 1]); hold on plot(scale*[d2(mkr, 1) p2(1)], scale*[d2(mkr, 2) p2(2)],`yo-`);endend%% Display all the Tracking Correspondencesfor i = 1:length(trackcor)if (trackcor{i},t == t) mkr = trackcor{i}.mkr; pul = trackcor{i}.pos(1,[2 1]); plr = trackcor{i}.pos(2,[2 1]); p2 = (pul+plr)/2; hold on plot(scale*[d2(mkr, 1) p2(1)], scale*[d2(mkr, 2) p2(2)],`yo-`); if (trackcor{i}.first == t) plot(scale*[pul(1) pul(1) plr(1) plr(1) pul(1)],scale*[pul(2) plr(2) plr(2) pul(2) pul(2)], `b+-`); else plot(scale*[pul(1) pul(1) plr(1) plr(1) pul(1)],scale*[pul(2) plr(2) plr(2) pul(2) pul(2)], `g+-`); endendendendif (length(frames) > 1), pause(0.1); endendfile: /homes/leventon/Matlab/showOverlay.mfunction showOverlay(seq, seqIm, ptcor, trackcor, frames, filespec)% showOverlay Show an Overlay of a reconstructed 3d sequence on an% image% sequence, with the point and tracking correspondences%% (See addConstraints -- this is similar except only a viewer)if (exists(seqIm, `scale`) == 0), seqIm.scale = 1; endscale = seqIm.scale / seq.scale;t = 1.; - n = size(seq.data, 1);d = seq.data(2:n, :);nFr = seq.nframes;global seqINFonmkr = seqINFO.nurnmkr(seq.subtype);d = reshape(d', 3, nmkr, nFr);d = transpose2(d);if (nargin < 4) trackcor = cell(0); endif (nargin < 3) ptcor = cell(0); endif (nargin < 6) filespec = [ ]; endif (nargin < 5) frames = 1:nFr; end% clffor t = frameshold offseqView2({seqIm, seq}, t);axis offd2 = worldToCam(d(:,:,t));for i = 1:length(ptcor)if (ptcor{i}.t == t)mkr = ptcor{i}.mkr;p2 = ptcor{i}.pos([2 1]);hold onplot(scale*[d2(mkr, 1) p2(1)], scale*[d2(mkr, 2) p2(2)], `yo-`);endendfor i = 1:length(trackcor)if (trackcor{ii}.t == t)mkr = trackcor{i}.mkr;pul = trackcor{i}.pos(1,[2 1]);plr = trackcor{i}.pos(2,[2 1]);p2 = (pul+plr)/2;hold onplot(scale*[d2(mkr, 1) p2(1)], scale*[d2(mkr, 2) p2(2)], `yo-`);if (trackcor{i}.first == t) plot(scale*[pul(1) pul(1) plr(1) plr(1) pul(1)], scale*[pul(2) plr(2) plr(2) pul(2) pul(2)], `b+-`);else plot(scale*[pul(1) pul(1) plr(1) plr(1) pul(1)], scale*[pul(2) plr(2) plr(2) pul(2) pul(2)], `g+-`);endendendif (length(frames)>1), pause(0.1); endif (isempty(filespec)==0)s = sprintf(filespec, t);print(`-dpsc`, s);endendfile: /homes/leventon/Matlab/showcurves.mfunction showcurves(A)n = size(A, 2);for i=4:23for j = 1:nwhile (A(i,j)<-pi)A(i,j) = A(i,j) + 2*pi;end;while (A(i,j)>pi)A(i,j) = A(i,j) - 2*pi;end;endendfigure(1);for i = 4:23hold off;subplot(5,4,i-3);;plot([0 1], [pi pi], `k`);hold onplot([0 1], [-pi -pi], `k`);plot([0 0], [-pi pi], `k`);plot([1 1], [-pi pi], `k`);;plot(linspace(0, 1, n), A(i,:));axis([0 1 -3.2 3.2]);axis offpause(0.1);end;subplot(5,4,1); title(`Body X`)subplot(5,4,2); title(`Body Y`)subplot(5,4,3); title(`Body Z`)subplot(5,4,4); title(`Torso Z`)subplot(5,4,5); title(`Rt Up Arm X`)subplot(5,4,6); title(`Rt Up Arm Y`)subplot(S,4,7); title(`Rt Lo Arm X`)subplot(5,4,8); title(`Rt Lo Arm Y`)subplot(5,4,9); title(`Lt Up Arm X`)subplot(5,4,10); title(`Lt Up Arm Y`)subplot(5,4,11); title(`Lt Lo Arm X`)subplot(S,4,12); title(`Lt Lo Arm Y`)subplot(5,4,13); title(`Rt Up Leg X`)subplot(5,4,14); title(`Rt Up Leg Y`)subplot(5,4,15); title(`Rt Lo Leg X`)subplot(5,4,16); title(`Rt Lo Leg Y`)subplot(5,4,17); title(`Lt Up Leg X`)subplot(5,4,18); title(`Lt Up Leg Y`)subplot(5,4,19); title(`Lt Lo Leg X`)subplot(5,4,20); title(`Lt Lo Leg Y`)file: /homes/leventon/Matlab/showit.mfunction showit(d, seq, is, ti, fn, smth)if (nargin < 3)is = 1:10;endif (nargin < 4)ti = {[ ], [ ]};endif (nargin < 5)fn = [ ];endif (nargin < 6)smth = 0;endn = size(d, 1);n = n / 10 / 4;n2 = n;= sqrt(n);for i = ist = (i-1)*n*n*4;subplot(1,2,2);I = [matricize(d(t+(1:n2)), n) matricize(d(n2+t+(1:n2)), n);matricize(d(2*n2+t+(1:n2)), n) matricize(d(3*n2+t+(1:n2)), n)];imagesc(smooth(I,smth));axis imagecaxis([-pi pi]);title(ti{2});subplot(1,2,1);seqView(seq, i, cons(`headalign`, 1, `snippetsize`, 10,`downsample`, 2, `smooth`, smth))title(ti{1});pause(0.1);if (isempty(fn) == 0)s = sprintf(fn, i);print(`-dpsc`, s);endendfile: /homes/leventon/Matlab/smed.m% InitializeseqInitWHICH.sub.-- SEQ = 1:10;WHICH.sub.-- NEV = [20 50 100 200];WHICH.sub.-- NOISE = [10 100 1000];WHICH.sub.-- SEQ = 1:10;WHICH.sub.-- NEV = [50 100];WHICH.sub.-- NOISE = 100;ERROR.sub.-- SHORT = zeros(4, length(WHICH.sub.-- SEQ),length(WHICH.sub.-- NEV));ERROR.sub.-- LONG = zeros(4, length(WHICH.sub.-- SEQ), length(WHICH.sub.-- NEV));ERROR.sub.-- KAL = zeros(4, length(WHICH.sub.-- SEQ), length(WHICH.sub.--NEV),length(WHICH.sub.-- NOISE));SEQS.sub.-- SHORT = cell(length(WHICH.sub.-- SEQ), length(WHICH.sub.--NEV));SEQS.sub.-- LONG = cell(length(WHICH.sub.-- SEQ), length(WHICH.sub.--NEV));SEQS.sub.-- KAL = cell(length(WHICH.sub.-- SEQ), length(WHICH.sub.--NEV),length(WHICH.sub.-- NOISE));for i.sub.-- sel.sub.-- index = 2:length(WHICH.sub.-- SEQ)sel.sub.-- index = WHICH.sub.-- SEQ(i.sub.-- sel.sub.-- index);for i.sub.-- nEv = 1:length(WHICH.sub.-- NEV)nEv = WHICH.sub.-- NEV(i.sub.-- nEv);%% Set numbers of E-vects% nEv = 50;disp(sprintf(`*** Leaving out sequence #%d`, sel.sub.-- index));% Train on all but one sequence% sel.sub.-- index = 3;%% all but sel.sub.-- index. Works for 1 and 10, too.train.sub.-- set = [1:(sel.sub.-- index-1) (sel.sub.-- index+1):10];sel = seqSelect(train.sub.-- set, pi/2);knl3d = seqTrain(sel, `3d`);knl2d = seqTrain(sel, `2d`);%% The encoded training matriciesM2 = knl2d.enc{1};M3 = knl3d.enc{1};%% Taking the SVD of these training matricies[U2, S2, V2] = svd(M2, 0);[U3, S3, V3] = svd(M3, 0);%% Get the observationobs2d = seqGet(sel.sub.-- index, `2d`, pi/2);obs3d = seqGet(sel.sub.-- index, `3d`, pi/2);% obs2dc = seqCrop(obs2d, 1, 22);% obs3dc = seqCrop(obs3d, 1, 22);obs2dc = obs2d;obs3dc = obs3d;%% Encode the observation (3 snippets worth)enc2dc = seqEncode(obs2dc);%% Let y be just the first snippet of the observationy = enc2dc(:,1);%% Make a projection matrix that takes a matrix%% from (3*370 .times. N) to (2*370 .times. N)P = project.sub.-- matrix(370);S3i = sparse(pinv(full(S3)));S2i = sparse(pinv(full(S2)));%% the true sequence% figure(1)% seqView(obs3dc);%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% This is the short path around the diagram (P * U3)% figure(2)[U, S, V] = svd(P * U3(:,1:nEv), 0);Si = sparse(pinv(full(S)));alpha = V * Si(:, 1:nEv) * U(:,1:nEv)' * enc2dc;fprintf(1, `short path fit to 2d data: %f.backslash.n`, . . . norm(enc2dc(:,1) - P * U3(:,1:nEv) * alpha(1:nEv, 1)));seqShort = seqDecode(U3(:,1:nEv) * alpha(1:nEv,:), `3d`, . . .kn12d.encParams, obs2d);% seqView(seqSshort, [ ]);ERROR.sub.-- SHORT(1, i.sub.-- sel.sub.-- index, i.sub.-- nEv)= seqError(seqShort, obs3d, `2d`);ERROR.sub.-- SHORT(2, i.sub.-- sel.sub.-- index, i.sub.-- nEv)= seqError(seqshort, obs3d, `3d`,0);ERROR.sub.-- SHORT(3, i.sub.-- sel.sub.-- index, i.sub.-- nEv)= seqError(seqShort, obs3d, `3d`,1);ERROR.sub.-- SHORT(4, i.sub.-- sel.sub.-- index, i.sub.-- nEv)= seqError(seqshort, obs3d, `3d`,2);SEQS.sub.-- SHORT{i.sub.-- sel.sub.-- index, i.sub.-- nEv} = seqShort;%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% This is the long path around the diagram (that works)%% figure(3)alpha.sub.-- star = 53(1:nEv,:) * V3' * V2 * S21(:,1:nEv) * U2(:, 1:nEv)'*enc2dc;fprintf(1, `long path fit to 2d data: %f.backslash.n`, . . . norm(enc2dc(:,1) - P * U3(:,1:nEv) * alpha.sub.-- star(1:nEv,1)));seqLong = seqDecode(U3(:,1:nEv) * alpha.sub.-- star(1:nEv,:), `3d`, . ..knl2d.encParams, obs2d);% seqView(seqLong, [ ]);ERROR.sub.-- LONG(1, i.sub.-- sel.sub.-- index, i.sub.-- nEv)= seqError(seqLong, obs3d, `2d`);ERROR.sub.-- LONG(2, i.sub.-- sel.sub.-- index, i.sub.-- nEv)= seqError(seqLong, obs3d, `3d`, 0);ERROR.sub.-- LONG(3, i.sub.-- sel.sub.-- index, i.sub.-- nEv)= seqError(seqLong, obs3d, `3d`, 1);ERROR.sub.-- LONG(4, i.sub.-- sel.sub.-- index, i.sub.-- nEv)= seqError(seqLong, obs3d, `3d`, 2);SEQS.sub.-- LONG{i.sub.-- sel.sub.-- index, i.sub.-- nEv} = seqShort;%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% Try the kalman filter MAP estimate of the 3d.%% (This should be very similar (identical?) to the "short path",%% with a prior term included of the proper weight).%% Work with mean zero training matrix. Just remember to add back%% in the mean at the end.mean3 = mean(M3, 2);M3zeromean = M3 - repmat(mean3, 1, size(M3,2));[U3O, S30, V30] = svd(M3zeromean, 0);%% do the kalman filter reconstruction (bayesian).%% From wtf composition book,for i.sub.-- obs.sub.-- noise = 1:length(WHICH.sub.-- NOISE)obs.sub.-- noise = WHICH.sub.-- NOISE(i.sub.-- obs.sub.-- noise);% obs.sub.-- noise = 100;g = S30(1:nEv,1:nEv) * U30(:,1:nEv)' * P' * . . . inv(P*U30(:,1:nEv)*S30(1:nEv,1:nEv)*U30(:,1:nEv))' * P' . . . + obs.sub.-- noise*eye(size(y,1)));%% modify the observations, enc2dc, to be zero mean, too.alphak = g * (enc2dc - repmat(P * mean3, 1, size(enc2dc,2)));% figure(4)seqKal = seqDecode(repmat(mean3,1,size(enc2dc,2)) + . . . U30(:,1:nEv) * alphak(1:nEv,:), `3d`, . . .knl2d.encParams, obs2d);% seqView(seqKal, [ ]);fprintf(1, `kalman fit to 2d data: %f.backslash.n`, . . . norm(enc2dc(:,1) - . . . P * (mean3 + U30(:,1:nEv) * alphak(1:nEv, 1))));KalError3d = seqError(seqKal, obs3d, `3d`, 1)ERROR.sub.-- KAL(1, i.sub.-- sel.sub.-- index, i.sub.-- nEv, i.sub.--obs.sub.-- noise) = seqError(seqKal,obs3d, `2d`);ERROR.sub.-- KAL(2, i.sub.-- sel.sub.-- index, i.sub.-- nEv, i.sub.--obs.sub.-- noise) = seqError(seqKal,obs3d, `3d`, 0);ERROR.sub.-- KAL(3, i.sub.-- sel.sub.-- index, i.sub.-- nEv, i.sub.--obs.sub.-- noise) = seqError(seqKal,obs3d, `3d`, 1);ERROR.sub.-- KAL(4, i.sub.-- sel.sub.-- index i.sub.-- nEv, i.sub.--obs.sub.-- noise) = seqError(seqKal,obs3d, `3d`, 2);SEQS.sub.-- KAL{i.sub.-- sel.sub.-- index, i.sub.-- nEv, i.sub.--obs.sub.-- noise} = seqKal;postCovDiag = diag(S30(1:nEv,1:nEv) - . . . g * P * U30(:,1:nEv) * S30(1:nEv,1:nEv));POST.sub.-- COV.sub.-- DIAG{i.sub.-- sel.sub.-- index, i.sub.-- nEv,i.sub.-- obs.sub.-- noise} = postCovDiag;endendendfile: /homes/leventonlMatlab/smooth.mfunction Id = smooth(I, t)% Id = smooth(I, t)% Does smoothing on the 2 or 3-d array, I.% Recursively smooths with a 1 2 1 filter, in each dimension.% Inefficient (should make a pyramid version of this).%% summer 97 mel wrote it.% 9/29/97 wtf wrote these comments.% 10/31/97 wtf empirically, I find that the smoothing factor t% is sigma 2 in a gaussian.if (nargin < 2)t = 1;endif (t < 0.5)Id = 1;returnendif (t > 1)I = smooth(I, t-1);end[m, n, p] = size(I);I = I + I([1 1:m-1], :, :);I = I(:, [1 1:n-1], :) + I;I = I + I([2:m m], :, :);I = I(:, [2:n n], :) + I;if (t < 1)Id = I/16;returnendI = I + I([1 1:m-1], :, :);I = I(:, [1 1:n-1], :) + I;I = I + I([2:m m], : , :)I = I(:, [2:n n], :) + I;% Id = I(1:2:m, 1:2:n, :)/256;Id = I/256;file: /homes/leventon/Matlab/spscaoa.mfunction A2 = spscaoa(A1, s)if (s == 1)A2 = A1;returnNO = 1endnfr = A1(1,1);n = nfr * 37;ndim = size(A1, 2);B1 = A1(2:n+1, :);B2 = s * B1;if (ndim == 2)A2 = [[nfr 37]; B2];elseA2 = [[nfr 37 30]; B2];endfile: /homes/leventon/Matlab/start.mload /homes/leventon/Snowman/dance01.mdtload /homes/leventon/Snowman/dance02.mdtload /homes/leventon/Snowman/dance03.mdtload /homes/leventon/Snowman/dance04.mdtload /homes/leventon/Snowman/destroy.mdtload /homes/leventon/Snowman/walkloop.mdtload /homes/leventon/Snowman/runloop.mdtload /homes/leventon/Snowman/fwalk.mdtload /homes/leventon/Snowman/frun.mdtload /homes/leventon/Snowrnan/thief.mdtdancec1 = [0 5; 5 6; 0 8; 8 9; 0 1; 0 4; 0 7; 4 7; 7 24; 24 25; 25 26;9 25; 6 22; 4 21; 21 22; 22 23; 10 11; 11 12; 12 13; 13 31; 27 28; 2829; 29 31; 30 31; 32 33; 14 15; 15 16; 16 17; 17 36; 33 34; 34 36; 3536; 3 19; 19 20; 20 3; 7 15; 15 33; 33 7; 4 28; 28 11; 11 4; 1 15; 111; 17 35; 13 30]load /homes/leventon/AOA/ballet.mdtfoo = ballet(2:9181,[3 2]);ballet(2:9181,[2 3]) = foo;load /homes/leventon/AOA/kicke.mdtfoo = kicke(2:4643, [3 2]);kicke(2:4643,[2 3]) = foo;if (0)clear ISIS{1} = dance01;IS{2} = dance02;IS{3} = dance03;IS{4} = dance04;IS{5} = destroy;IS{6} = walkloop;IS{7} = runloop;IS{8} = fwalk;IS{9} = frun;IS{10} = thief;endnmk = 37;% M1 = mencode(IS{1}, IS{2}, IS{3}, IS{4}, IS{5}, IS{6}, IS{7}, IS{8},% IS{9}, IS{10}, [10 5]);% [u1, s1, v1, E1] = decomp(M1, nmk);if (1)clear ISaISa{1} = mkr2ang(dance01);ISa{2} = mkr2ang(dance02);ISa{3} = mkr2ang(dance03);ISa{4} = mkr2ang(dance04);ISa{5} = mkr2ang(destroy);ISa{6} = mkr2ang(walkloop);ISa{7} = mkr2ang(runloop);ISa{8} = mkr2ang(fwalk);ISa{9} = mkr2ang(frun);ISa{10} = mkr2ang(thief);person = mcylmodel(walkloop);endif (0)dance01r45 = rotaoa(dance01, pi/4);dance02r45 = rotaoa(dance02, pi/4);dance03r45 = rotaoa(dance03, pi/4);dance04r45 = rotaoa(dance04, pi/4);destroyr4s = rotaoa(destroy, pi/4);walkloopr45 = rotaoa(walkloop, pi/4);runloopr45 = rotaoa(runloop, pi/4);fwalkr45 = rotaoa(fwalk, pi/4);frunr45 = rotaoa(frun, pi/4);thiefr45 = rotaoa(thief, pi/4);dance01r90 = rotaoa(dance01, pi/2);dance02r90 = rotaoa(dance02, pi/2);dance03r90 = rotaoa(dance03, pi/2);dance04r90 = rotaoa(dance04, pi/2);destroyr90 = rotaoa(destroy, pi/2);walkloopr90 = rotaoa(walkloop, pi/2);runloopr90 = rotaoa(runloop, pi/2);fwalkr90 = rotaoa(fwalk, pi/2);frunr90 = rotaoa(frun, pi/2);thiefr90 = rotaoa(thief, pi/2);dance01p00 = proj2d(dance01);dance02p00 = proj2d(dance02);dance03p00 = proj2d(dance03);dance04p00 = proj2d(dance04);destroyp00 = proj2d(destroy);walkloopp00 = proj2d(walkloop);runloopp00 = proj2d(ninloop);fwalkp00 = proj2d(fwaik);frunp00 = proj2d(frun);thiefp00 proj2d(thief);dance01p4s = proj2d(dance01r45);dance02p45 = proj2d(dance02r45);dance03p45 = proj2d(dance03r4s);dance04p45 = proj2d(dance04r45);destroyp4s = proj2d(destroyr4s);walkloopp45 = proj2d(walkloopr45);runloopp45 = proj2d(runloopr4s);fwalkp45 = proj2d(fwalkr4s);frunp4s = proj2d(frunr4s);thiefp4s = proj2d(thiefr4s);dance01p90 = proj2d(dance01r90);dance02p90 = proj2d(dance02r90);dance03p90 = proj2d(dance03r90);dance04p90 = proj2d(dance04r90);destroyp90 = proj2d(destroyr90);walkloopp90 = proj2d(walkloopr90);runloopp90 = proj2d(runloopr90);fwalkp90 = proj2d(fwalkr90);frunp90 = proj2d(frunr90),thiefp90 = proj2d(thiefr90);endIS2 = cell(10,3);IS3 = cell(10,8);theta = 0;for i=1:8IS3{1,i} = rotaoa(dance01, theta);IS3{2,i} = rotaoa(dance02, theta);IS3{3,i} = rotaoa(dance03, theta);IS3{4,i} = rotaoa(dance04, theta);IS3{5,i} = rotaoa(destroy, theta);IS3{6,i} = rotaoa(walkloop, theta);1I3{7,i} = rotaoa(runloop, theta);IS3{8,i} = rotaoa(fwalk, theta);IS3{9,i} = rotaoa(frun, theta);IS3{10,i} = rotaoa(thief, theta);for j = 1:10I52{j,i} = proj2d(IS3{j,i});endtheta = theta + pi/4;endfor i = 1:10AoaSizes(i) = IS3{i}(1,1);endglobal tgf1global tgf2global tgf3global thf1global thf2global thf3global thf4tgf1 = [0.0094 0.1148 0.3964 -0.0601];tgf1 = [tgf1 -0.9213 tgfi(4:-1:1)];tgf2 = [0.0008 0.0176 0.1660 0.6383];tgf2 = [tgf2 1.0 tgf2(4:-1:1)];tgf3 = [0.0028 0.0480 0.3020 0.5806];tgf3 = [tgf3 0.0 -tgf3(4:-1:1)];thf1 = [0.0098 0.0618 -0.0998 -0.7551];thf1 = [thf1 0.0 -thfi(4:-1:1)];thf2 = [0.0008 0.0176 0.1660 0.6383];thf2 = [thf2 1.0 thf2(4:'11:1)];thf3 = [0.0020 0.0354 0.2225 0.4277];thf3 = [thf3 0.0 -thf3(4:-1:1)];thf4 = [0.0048 0.0566 0.1695 -0.1889];thf4 = [thf4 -0.7349 thf4(4:-1:1)];path(`/homes/leventon/matlabpyrTools', path);file: /homes/leventon/Matlab/stickify.mfunction I = stickify(seq, dnsamp)% I = stickify(seq, dnsamp)% Makes an image sequence of a stick figure of the body.%% summer, 1997 mel wrote it% 9/29/97 wtf wrote this comment.if (nargin < 2)dnsamp = 0;endglobal seqINFOif (isSequence(seq))%% It's a sequenceswitch seq.typecase `3d`seq = seqConvert(seq, `2d`);case `2d`otherwiseerror(`Only can stickify 2d or 3d seqences');endn = size(seq.data, 1);d = seq.data(2:n, :);N = seq.nframes;nmk = seqINFO.nummkr(seq.subtype);imsize = [90 160];dc = worldToCam(d);elseif (size(seq, 2) == 1)%% It's a snippetN = 10;nmk = 37;d = reshape(seq, 3, nmk * N)';imsize = [80 80];dc = worldToCam(d) + ones(nmk*N, 1) * ([40 40] - worldToCam([0 00]));endoris = [0 pi/4 pi/2 3*pi/4];nOris = length(oris);imsize = imsize / (2 dnsamp);for t = 1:NI{t} = zeros(imsize(1), imsize(2), nOris);c = seqINFO.connections{1};p1 = dc((t-1)*nmk + c(:,1) + 1, :);p2 = dc((t-1)*nmk + c(:,2) + 1, :);p1 = p1 / (2 dnsamp);p2 = p2 / (2 dnsamp);np = (p2-p1);np = np ./ (sqrt(sum(np. 2, 2)) * ones(1,2));thetas = ((np * [cos(oris); sin(oris)])`)`;val = cos(2*thetas). 2;I{t} = renderLine2(I{t}, p1(:,[2 1]), p2(:,[2 1]), val);endfile: /homes/leventon/Matlab/stickify0.mfunction I = stickify(seq)global seqINFOswitch seq.typecase `3d`seq = seqConvert(seq, `2d`);case `2d`otherwiseerror(`Only can stickify 2d or 3d seqences`);endoris = [0 pi/4 pi/2 3*pi/4]nOris = length(oris);imsize = [90 160];n = size(seq.data, 1);d = seq.data(2:n, :);dc = worldToCam(d);nmk = 37;for t = 1:seq.nframesI{t} = zeros(imsize(1), imsize(2), nOris);for c = seqINFO.connections{1}'% p1 = dc((t-1)*nmk + c(:,1) + 1, :);% p2 = dc((t-1)*nmk + c(:,2) + 1, :);% np = (p2-p1). 2;% np = sqrt(sum(np, 2));p1 = dc((t-)*nmk + c(i) + 1, :);p2 = dc((t-)*nmk + c(2) + 1, :);np = normalize(p2-p1)for i.sub.-- o = 1:nOristheta = oris(i.sub.-- o);theta = dot([cos(theta) sin(theta)], np);val = cos(2*theta) 2;J = I{t}(:,:,i.sub.-- o);J = renderLine2(J, p1([2 1]), p2([2 1]), val);I{t}(:,:,i.sub.-- o) = J;endendendfile: /homes/leventon/Matlab/stickify2.mfunction I = stickify(seq)global seqINFOif (isSequence(seq))%% It's a sequenceswitch seq.typecase `3d`seq = seqconvert(seq, `2d`);case `2d`otherwiseerror(`Only can stickify 2d or 3d seqences`);endn = size(seq.data, 1);d = seq.data(2:n, :);N = seq.nframes;nmk = seqINFO.nummkr(seq.subtype);imsize = [90 160];dc = worldToCam(d);elseif (size(seq, 2) == 1)%% It's a snippetN = 1;nmk = 37;d = reshape(seq, 3, nmk * N)';imsize = [80 80];dc = worldToCam(d) + ones(nmk*N, 1) * ([40 40] - worldToCam([0 00]));endoris = [0 pi/4 pi/2 3*pi/4];nOris = length(oris);for t = 1:1I{t} = zeros(imsize(1), imsize(2), nOris);c = seqINFO.connections{1};p1 = dc((t-)*nmk + c(:,1) + 1, :);p2 = dc((t-)*nmk + c(:,2) + 1, :);np = (p2-p1);np = np ./ (sqrt(sum(np. 2, 2)) * ones(1,2));thetas = ((np * [cos(oris); sin(oris)])`)`;val = cos(2*thetas). 2;I{t} = gausline(I{t}, p1(:,[2 1]), p2(:,[2 1]), val);endfile: /homes/leventon/Matlab/stop.sub.-- ani.sub.-- fig.mWHICH.sub.-- SEQ = 4;WHICH.sub.-- FRAMES = 50:5:95;WHAT.sub.-- ANG = pi/2;% WHICH.sub.-- SEQ = 6;% WHICH.sub.-- FRAMES = 1:7:46;% WHAT.sub.-- ANG = pi/2;NUM.sub.-- EVS = [10 20 40];NUM.sub.-- EVS.sub. 2D = 40;n = 1:10;sel = seqselect(n(find(n .sup..about. = WHICH.sub.-- SEQ)), WHAT.sub.--ANG);knl3d = seqTrain(sel, `3d`);obs3d = seqGet(WHICH.sub.-- SEQ, `3d`, WHAT.sub.-- ANG);for nev = NUM.sub.-- EVSrcn = seqRecon(obs3d, knl3d, `3d`, nev);seqView(rcn, WHICH.sub.-- FRAMES, `stopani`, 1);endknl2d = seqTrain(sel, `2d`);obs2d = seqGet(WHICH.sub.-- SEQ, `2d`, WHAT.sub.-- ANG);rcn2d = seqRecon(obs2d, knl2d, `3d`, NUM.sub.-- EVS.sub.-- 2D);globa.sub.-- hack.sub.-- override.sub.-- viewaoa = [ ];seqView(rcn, WHICH.sub.-- FRAMES, `stopani`, 1);axis( [-121.9112 82.3015 -75.5499 33.1390 -5.5572 104.2409])colormap(gray(50))c = colormap;colormap(c(50:-1:1, :));file: /homes/leventon/Matlab/subbg.mfunction FG = subbg(IM, BG)[m, n, N] = size(IM);for i = 1:NSTATUS= iId = imDiff(IM(:,:,i)-median(vectorize(IM(:,:,i))),BG-median(vectorize(BG)), 10);Id = Id / max(max(Id));Ithresh = Id>0.10;Ierode = erode(Ithresh, ones(3));Idilate = double(dilate(Ierode, circle(6)));Ismooth(:,:,i) = smooth(Idilate,8);endFG = reshape(Ismooth, m*n, N);FG = conv2(1, [1 4 6 4 1], FG);FG = FG(:,3:N+2);FG = reshape(FG, m, n, N);FG = FG .* IM;file: /homes/leventon/Matlab/superimp.mfunction I = superimp(model, im, trans, star)if (nargin < 4)star = 0;endIm = zeros(Lsize(im) 3);Im(:,:,1) = double(im);Im(:,:,2) = double(im);Im(:,:,3) = double(im);if (iscell(model) == 0)model = {model};trans = {trans};endfor i = 1:length(model)if (isempty(trans{i}) == 0)w = size(model{i}, 1);h = size(model{i}, 2);trans{i} = trans{i}';xs = trans{i}(1):trans{i}(1)+w-1;ys = trans{i}(2):trans{i}(2)+h-1;Im(xs, ys, 1) = max(Im(xs, ys, 1), double((model{i}.sup..about. =1)));Im(xs, ys, 2) = min(Im(xs, ys, 2), double((model{i}==1)));endendif (nargout == 1)I = Im;elseimagesc(Im);axis irnage;hold onfor i = 1:length(model)if (isempty(trans{i}) == 0)w = size(model{i}, 1);h = size(model{i}, 2);xm = trans{i}(1) + w/2;ym = trans{i}(2) + h/2;plot(ym, xm, `b*`);endendhold offaxis offendfile: /homes/leventon/Matlab/svdinv.mfunction Ai = svdinv(A, nEv)if (nargin < 2)nEv = min(size(A));endnEv = min(nEv, min(size(A)));[U, S, V] = svd(A, 0);Si = spdiags(1./diag(S), 0, size(V,1), nEv);Ai = V*Si*U(:,1:nEv));file: /homes/leventon/Matlab/temp.mfunction h2 = temp(h)h2 = zeros(size(h));h2(:,1) = 160-((h(:,1)-43.19)*(55-125)/(52.19+87.9) + 57);h2(:,2) = (h(:,2)-43.19)*(55-125)/(52.19+87.9) + 57;h2(:,3) = (h(:,3)-40)*(45-30)/(40-70) + 44;file: /homes/leventon/Matlab/test.sub.-- script.m% InitializeseqInit% Train on all but one sequencesel.sub.-- index = 3;%% all but sel.sub.-- index. Works for 1 and 10., too.train.sub.-- set = [1:(sel.sub.-- index-1) (sel.sub.-- index+1):10];sel = seqSelect(train.sub.-- set, pi/2);knl3d = seqTrain(sel, `3d`);knl2d = seqTrain(sel, `2d`);%% The encoded training matriciesM2 = knl2d.enc{1};M3 = knl3d.enc{1};%% Taking the SVD of these training matricies[U2, S2, V2] = svd(M2, 0);[U3, S3, V3] = svd(M3, 0);%% Get the observationobs2d = seqGet(sel.sub.-- index, `2d`, pi/2);obs3d = seqGet(sel.sub.-- index, `3d`, pi/2);obs2dc = seqcrop(obs2d, 1, 22);obs3dc = seqcrop(obs3d, 1, 22);%% Encode the observation (3 snippets worth)enc2dc = seqEncode(obs2dc);%% Let y be just the first snippet of the observationy = enc2dc(:,1);%% Make a projection matrix that takes a matrix%% from (3*370 .times. N) to (2*370 .times. N)P = project.sub.-- matrix(370);%% Set numbers of E-vectsnEv = 50;S3i = sparse(pinv(full(S3)));S2i = sparse(pinv(full(S2)));%% the true sequencefigure (1)seqView(obs3dc);%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% This is the short path around the diagram (P * U3)figure(2)[U, S, V] = svd(P * U3(:,1:nEv), 0);Si = sparse(pinv(full(S)));alpha = V * Si(:, 1:nEv) * U(:,1:nEv)' * enc2dc;fprintf(1, `short path fit to 2d data: %f.backslash.n`, . . . norm(enc2dc(:,1) - P * U3(:,1:nEv) * alpha(1:nEv, 1)));seqView(seqDecode(U3(:,1:nEv) * alpha(1:nEv,:), `3d`, . . .knl2d.encParams, obs2d), [ ])%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% This is the long path around the diagram (that works)figure (3)alpha.sub.-- star = S3(1:nEv,:) * V3' * V2 * S21(:,1:nEv) * U2(:, 1:nEv)'*enc2dc;fprintf(1, `long path fit to 2d data: %.f.backslash.n`, . . . norm(enc2dc(:,1) - P * U3(:,1:nEv) * alpha.sub.-- star(1:nEv,1)));seqView(seqDecode(U3(:,1:nEv) * alpha.sub.-- star(1:nEv,:), `3d`, . . .knl2d.encParams, obs2d), [ ])%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% Try the kalman filter MAP estimate of the 3d.%% (This should be very similar (identical?) to the "short path",%% with a prior term included of the proper weight).%% Work with mean zero training matrix. Just remember to add back%% in the mean at the end.mean3 = mean(M3, 2);M3zeromean = M3 - repmat(mean3, 1, size(M3,2));[U30, S30, V30] = svd(M3zeromean, 0);%% do the kalman filter reconstruction (bayesian).%% From wtf composition book,obs.sub.-- noise = 100;g = S30(1:nEv,1:nEv) * U30(:,1:nEv)' * P' * . . . inv(P*U30(:,1:nEv)*S30(1:nEv,1:nEv)*U30(:,1:nEv))'* P' . . . + obs.sub.-- noise*eye(size(y,1)));%% modify the observations, enc2dc, to be zero mean, too.alphak = g * (enc2dc - repmat(P * mean3, 1, size(enc2dc,2)));figure (4)seqView(seqDecode(repmat(mean3,1,size(enc2dc,2)) + . . .U30(:,1:nEv) * alphak(1:nEv,:); `3d`, . . .knl2d.encParams, obs2d), [])fprintf(1, `kalman fit to 2d data: %f.backslash.n`, . . . norm(enc2dc(:,1) - . . . P * (mean3 + U30(:,1:nEv) * alphak(1:nEv, 1))));%% The covariance matrix of the posterior distribution of the%% 3-d figure. From the standard bayesian/kalman filter completing%% the square. This posterior covariance should tell you how much%% uncertainty there is about the figure after the measurement, and%% in what directions it is.postCov = S30(1:nEv,1:nEv) - . . . g * P * U30(:,1:nEv) * S30(1:nEv,1:nEv);figure (5)semilogy(diag(S30), `r`);v = axis; v(2) = 50; v(3) = 10; v(4) = 100000;hold on;semilogy(diag(postCov), `g`);axis(v);hold off;title(`prior: r posterior: g`);%% Dominated by contribution in mode U30(:,2)%% Display that mode: (or else the next 2 highest: 1 and 10).figure (6)subplot(3,1,1)%% we don't know how he's sliding;seqView(seqDecode(-10 * U30(:,2), `3d`, . . .knl2d.encParams, obs2d), [])subplot(3,1,2)%% we don't know where he's standingseqView(seqDecode(10 * U30(:,1), `3d`, . . .knl2d.encParams, obs2d), [ ])subplot(3,1,3)%% we don't know how he's moving his arms towards the cameraseqView(seqDecode(U30(:,10) , `3d`, . . .knl2d.encParams, obs2d), [ ])%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%return;%% redisplay all the resultsfigure(1)seqView(obs3dc);figure (2)seqView(seqDecode(U3(:,1:nEv) * alpha(1:nEv,:), `3d`, . . .knl2d.encParams, obs2d), [ ])figure(3)seqView(seqDecode(U3(:,1:nEv) * alpha.sub.-- star(1:nEv,:), `3d`, . . .knl2d.encParams, obs2d), [ ])figure (4)seqView(seqDecode(repmat(mean3,1,size(enc2dc,2)) + . . . U30(:,1:nEv) * alphak(1:nEv,:), `3d`, kn12d.encParams, obs2d), [ ]) return %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% %% %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% %% %%% This is an attempt to add the alpha' * S2i * alpha constraint %% to the short path around the diagram lambda = 1; [Up, Sp, Vp] = svd(U3' * P' * P * U3 - lambda * S2i, 0); Spi = sparse(pinv(full(Sp))); alpha = Vp * Spi(:,1:nEv) * Up(:,1:nEv)' * U3' * P' * enc2dc; seqView(seqDecode(U3(:,1:nEv) * alpha(1:nEv,:), . . . `3d`, knl2d.encParams, obs2d), []);%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% This is an attempt to add the alpha' * S2i * alpha constraint%% to the long path around the diagramB = U2(:, 1:nEv) * S2(1:nEv, :) * V2' * V3 * S3i;BiStar = S3 * V3' * V2 * S2i(:,1:nEv) * U2(:, 1:nEv)';%% Solve: (B'*B - S3i) * alpha = B'*y%% (*) Bill, I just found a bug in my derivation: The line%% above and below this are not equivalent. I should not%% have factored out a B' from the - S3i . . .%% Solve: B' * (U2 * S2 * V2' * V3 - I) * S3i * alpha = ' * y%% Solve: (U2 * S2 * V2' * V3 - I) * S3i * alpha = y%% Solve: (S2 * V2' * V3 - I) * S3i * alpha = U2' * y%% Solve: S3i * alpha = pinv(S2 * V2' * V3 - I) * U2' * y%% Solve:alpha = S3 * pinv(S2 * V2' * V3 - I) * U2' * y%% However, what's below seems to work? ? (even given the bug (*))lambda = 10;S2.sub.-- V2t.sub.-- V3.sub.-- m.sub.-- U2t.sub.-- pinv = pinv((S2(1:nEv,:) * V2' * V3) - . . .lambda * U2(:, 1:nEv)' * eye(740, 294));alpha = (S3 * S2.sub.-- V2t.sub.-- V3.sub.-- m.sub.-- U2t.sub.-- pinv *U2(:,1:nEv)' * enc2dc);seqView(seqDecode(U3(: ,1:nEv) * alpha(1:nEv,:), `3d`, . . .knl2d.encParams, obs2d), [ ])%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%file: /homes/leventon/Matlab/test.sub.-- script.sub.-- loop.m% InitializeseqInitWHICH.sub.-- SEQ = 1:10;WHICH.sub.-- NEV = [20 50 100 200];WHICH.sub.-- NOISE = [10 100 1000];WHICH.sub.-- SEQ = 9;WHICH.sub.-- NEV = 50;WHICH.sub.-- NOISE = 100;ERROR.sub.-- SHORT = zeros(4, length(WHICH.sub.-- SEQ),length(WHICH.sub.-- NEV));ERROR.sub.-- LONG = zeros(4, length(WHICH.sub.-- SEQ), length(WHICH.sub.-- NEV));ERROR.sub.-- KAL = zeros(4, length(WHICH.sub.-- SEQ), length(WHICH.sub.--NEV),length(WHICH.sub.-- NOISE));for i.sub.-- sel.sub.-- index = 1:length(WHICH.sub.-- SEQ)sel.sub.-- index = WHICH.sub.-- SEQ(i.sub.-- sel.sub.-- index);for i.sub.-- nEv = 1:length(WHICH.sub.-- NEV)nEv = WHICH.sub.-- NEV(i.sub.-- nEv);*%% Set numbers of E-vects% nEv = 50;disp(sprintf(`*** Leaving out sequence #%d`, sel.sub.-- index));% Train on all but one sequence% sel.sub.-- index = 3;%% all but sel.sub.-- index. Works for 1 and 10, too.train.sub.-- set = [1: (sel.sub.-- index-1) (sel.sub.-- index+1):10];sel = seqSelect(train.sub.-- set, pi/2);knl3d = seqTrain(sel, `3d`);knl2d = seqTrain(sel, `2d`);%% The encoded training matriciesM2 = knl2d.enc{1};M3 = knl3d.enc{1};% Taking the SVD of these training matricies[U2, S2, V2] = svd(M2, 0);[U3, S3, V3] = svd(M3, 0);%% Get the observationobs2d = seqGet(sel.sub.-- index, `2d`, pi/2);obs3d = seqGet(sel.sub.-- index, `3d`, pi/2);% obs2dc = seqcrop(obs2d, 1, 22);% obs3dc = seqcrop(obs3d, 1, 22);obs2dc = obs2d;obs3dc = obs3d;%% Encode the observation (3 snippets worth)enc2dc = seqEncode(obs2dc);%% Let y be just the first snippet of the observationy = enc2dc(:,1);%% Make a projection matrix that takes a matrix%% from (3*370 .times. N) to (2*370 .times. N)P = project.sub.-- matrix(370);S3i = sparse(pinv(full(S3)));S2i = sparse(pinv(full(S2)));%% the true sequencefigure(1)% seqView(obs3dc);%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% This is the short path around the diagram (P * U3)figure(2)[U, S, V] = svd(P * U3(:,1:nEv), 0);Si = sparse(pinv(full(S)));alpha V * Si(:, 1:nEv) * U(:,1:nEv)' * enc2dc;fprintf(1, `short path fit to 2d data: %f.backslash.n`, . . . norm(enc2dc(:,1) - P * U3(:,1:nEv) * alpha(1:nEv, 1)));seqShort = seqDecode(U3(:,1:nEv) * alpha(1:nEv,:), `3d`, . . .knl2d.encParams, obs2d);% seqView(seqShort, [ ]);ERROR.sub.-- SHORT(1, i.sub.-- sel.sub.-- index, i.sub.-- nEv)= seqError(seqShort, obs3d, `2d`);ERROR.sub.-- SHORT(2, i.sub.-- sel.sub.-- index, i.sub.-- nEv)= seqError(seqShort, obs3d, `3d`,0);ERROR.sub.-- SHORT(3, i.sub.-- sel.sub.-- index, i.sub.-- nEv)= seqError(seqShort, obs3d, `3d`,1);ERROR.sub.-- SHORT(4, i.sub.-- sel.sub.-- index, i.sub.-- nEv)= seqError(seqShort, obs3d, `3d`,2);%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% This is the long path around the diagram (that works)figure (3)alpha.sub.-- star = S3(1:nEv,:) * V3' * V2 * S2i(:,1:nEv) * U2(:, 1:nEv)'*enc2dc;fprintf(1, `long path fit to 2d data: %f.backslash.n`, . . . norm(enc2dc(:,1) - P * U3(:,1:nEv) * alpha.sub.-- star(1:nEv,1)));seqLong = seqDecode(U3(:,1:nEv) * alpha.sub.-- star(1:nEv,:), `3d`, . ..knl2d.encParams, obs2d);% seqView(seqLong, [ ]);ERROR.sub.-- LONG(1, i.sub.-- sel.sub.-- index, i.sub.-- nEv)= seqError(seqLong, obs3d, `2d`);ERROR.sub.-- LONG(2, i.sub.-- sel.sub.-- index, i.sub.-- nEv)= seqError(seqLong, obs3d, `3d`, 0);ERROR.sub.-- LONG(3, i.sub.-- sel.sub.-- index, i.sub.-- nEv)= seqError(seqLong, obs3d, `3d`, 1);ERROR.sub.-- LONG(4, i.sub.-- sel.sub.-- index, i.sub.-- nEv)= seqError(seqLong, obs3d, `3d`, 2);%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% Try the kalman filter MAP estimate of the 3d.%% (This should be very similar (identical?) to the "short path",%% with a prior term included of the proper weight).%% Work with mean zero training matrix. Just remember to add back%% in the mean at the end.mean3 = rnean(M3, 2);M3zeromean = M3 - repmat(mean3, 1, size(M3,2));[U30, S30, V30] = svd(M3zeromean, 0);%% do the kalman filter reconstruction (bayesian)%% From wtf composition book,for i.sub.-- obs.sub.-- noise = 1:length(WHICH.sub.-- NOISE)obs.sub.-- noise = WHICH.sub.-- NOISE(i.sub.-- obs.sub.-- noise);% obs.sub.-- noise = 100;g = S30(1:nEv,1:nEv) * U30(:,1:nEv)' * P' * . . . inv(P*U30(:,1:nEv)*S30(1:nEv,1:nEv)*U30(:,1:nEv)'* P' . . . + obs.sub.-- noise*eye(size(y,1)));%% modify the observations, enc2dc, to be zero mean, too.alphak = g * (enc2dc - repmat(P * mean3, 1, size(enc2dc,2)));figure (4)seqKal = seqDecode(repmat(mean3,1,size(enc2dc,2)) + . . .U30(:,1:nEv) * alphak(1:nEv,:), `3d`, . . .knl2d.encParams, obs2d);% seqView(seqKal, [ ]);fprintf(1, `kalman fit to 2d data: %f.backslash.n', . . . norm(enc2dc(:,1) - . . . P * (mean3 + U30(:,1:nEv) * alphak(1:nEv, 1))));KalError3d = seqError(seqKal, obs3d, `3d`, 1)ERROR.sub.-- KAL(1, i.sub.-- sel.sub.-- index, i.sub.-- nEv, i.sub.--obs.sub.-- noise) = seqError(seqKal,obs3d, `2d`);ERROR.sub.-- KAL(2, i.sub.-- sel.sub.-- index, i.sub.-- nEv, i.sub.--obs.sub.-- noise) = seqError(seqKal,obs3d, `3d`, 0);ERROR.sub.-- KAL(3, i.sub.-- sel.sub.-- index, i.sub.-- nEv, i.sub.--obs.sub.-- noise) = seqError(seqKal,obs3d, `3d`, 1);ERROR.sub.-- KAL(4, i.sub.-- sel.sub.-- index, i.sub.-- nEv, i.sub.--obs.sub.-- noise) = seqError(seqKal,obs3d, `3d`, 2);endendendfile: /homes/leventon/Matlab/timescaoa.mfunction A2 = timescaoa(A1, s)if (s == 1)A2 = A1;returnNO = 1endnfr = A1(1,1);n = nfr * 37;ndim = size(A1, 2);nfr2 = size(1:s:nfr, 2);n2 = nfr2 * 37;B1 = A1(2:n+1, :);B2 = zeros(n2, ndim);i = 1;for x = 1:s:nfrri = ((i-1)*37+1):(i*37);xf = floor(x);rf = ((xf-1)*37+1):(xf*37);xc = cell(x);rc = ((xc-1)*37+1):(xc*37);if (xf == xc)B2(ri, :) = B1(rc, :);elsedf = x - xf;dc = xc - x;B2(ri, :) = (dc * B1(rf, :)) + (df * B1(rc, :));endi = i + 1;endif (ndim == 2)A2 = [[nfr2 37]; B2];elseA2 = [[nfr2 37 30]; B2];endfile: /homes/leventon/Matlab/tmpstart.mseqInitinitSensorssel = seqselect(1:10, pi/2);knl3d = seqTrain(sel, `3d`, `MtM`, 0, `submean`, 1, `headalign`, 1);load . . ./Data/barishpts.matload . . ./Data/barish.matnEv = 50;S3i = spdiags(1./diag(knl3d.S), 0, size(knl3d.V,1), nEv);seq = seqcrop(seqGet(10), 1, seqSrn.nframes);seq.scale = seqsm.scale;enc = seqEncode(seq, `force`, 1);betai = knl3d.V * S3i(:,1:nEv) * knl3d.U(:,1:nEv)' * enc;betai0 = betai;alpha3 = knl3d.S(1:nEv,:) * knl3d.V' * betai;file: /homes/leventon/Matlab/tmpstart2.mseqInitsel = seqSelect(1:10, pi);knl3d = seqTrain(sel, `3d`, `MtM`, 0, `submean`, 1, `headalign`, 1);load . . ./Data/wiz.matif (0)load . . . /Data/dorthypts.matptcorr = ptcorr.sub.-- dorthy;else[seq0, alplia0, beta0, ptcorr] = seqFit(seqSm, `knl3d`, knl3d, . . .`snips`, 1:18, `fminiter`, 1000, `corrLambda`, 0.5, . . .`priorLambda`, 0.2, `ISSDELambda`, 0 `LenDELambda`, 0, . . .`nEv`, 50, `interactive`, 1);end[seq1, alpha1, beta1, ptcorr] = seqFit(seqSm, `knl3d`, knl3d, . . .`snips`, 1:18, `ptcorr`, ptcorr, `fminiter`, 2000, . . .`corrLambda`, 0.5, `priorLambda`, 0.2, `ISSDELambda`, 0, . . .`LenDELambda`, 0, `nEv`, 25, `interactive`, 0);[seq2, alpha2, beta2, ptcorr] = seqFit(seqSm, `knl3d`, knl3d, . . .`snips`, 1:18, `ptcorr`, ptcorr, `fminiter`, 1000, . . .`corrLambda`, 0.5, `priorLambda`, 0.2, `ISSDELambda`, 0.05,. . .`LenDELambda`, 0, `nEv`, 25, `interactive`, 0, `beta`, beta1);[seq3, alpha3, beta3, ptcorr] = seqFit(seqSm, `knl3d`, knl3d, . . .`snips`, 1:18, `ptcorr`, ptcorr, `fminiter`, 1000, . . .`corrLambda`, 0.5, `priorLambda`, 0.2, `ISSDELambrda`, 0.05,. . .`LenDELambda`, 0, `nEv`, 50, `interactive`, 0, `beta`, beta2);file: /homes/leventon/Matlab/tp.m%% tp.m test prior%% Check out different priors for the coefficients, to see if%% you can pose an optimization problem which gives the particular%% reconstruction that we want, in cases of overdetermined%% reconstructions. Just use a few training vectors. Not lots.%% Find the solution to the energy function,%% energy = sum((y - M * x). 2) + lambda * sum(abs(x). alpha);%% July 16, 1997 freeman created, starting from MEL posed problem.%% to pass to optimization callsglobal y M lambda alpha;n.sub.-- col = 10; n.sub.-- row = 10;%% make a rank deficient matrix. Let w be the last column.M = rand(n.sub.-- row, n.sub.-- col);M = [M sum(M`)`/n.sub.-- col];w = M(:, n.sub.-- col+1);[U, S, V] = svd(M);beta1 = V * pinv(S) * U' * w;beta2 = 1/(n.sub.-- col) * ones(n.sub.-- col+1, 1); beta2(n.sub.--col+1,1) = 0;beta3 = zeros(n.sub.-- col+1, 1); beta3(n.sub.-- col+1,1) = 1;val1 = M * beta1 - w;val2 = M * beta2 - w;val3 = M * beta3 - w;%% The energy function we want to optimize:lambda = 0.01; alpha = 0.5; y = w;disp(sprintf(`energy for lambda=%f and alpha=%f: soln1 %f, soln2 %f,%soln3 %,f`, . . .lambda, alpha, tp.sub.-- energy(beta1), tp.sub.-- energy(beta2),tp.sub.-- energy (beta3)));%% set-up for optimizationoptions(1) = 0; options(2) = 1e-8; options(3) = 1e-8; options(14) =50000;x0 = rand(n.sub.-- col+1, 1);xmin = x0;%% for frprmn optimization:%% ftol = 1e-12;figure(1); i.sub.-- plot = 0; j.sub.-- plot = 0;lambda.sub.-- list = [0 0.01 0.1 1];alpha.sub.-- list = [2 1 0.5 0.25 0.12];for lambda = lambda.sub.-- listfit = [ ];for alpha = alpha.sub.-- listi.sub.-- plot = i.sub.-- plot + 1;xmin = fmins(`tp.sub.-- energy`, xmin, options, `tp.sub.-- grad`);%%% For some reason, frprmn gives consistantly poorer results thanfmins,%%% for cases where alpha < 2. Why is that?%%% [xmin, cost, iter] = frprmn(xmin, ftol, `tp.sub.-- energy`,%%% `tp.sub.-- grad`);figure(1);subplot(size(lambda.sub.-- list, 2), size(alpha.sub.-- list, 2), i.sub.--plot);plot(xmin, `r*`);title([`wgt: ` num2str(lambda) `exp: ` num2str(alpha)]);axis([1, n.sub.-- col+1, -0.1, 1.1]);tmp.sub.-- lambda = lambda; lambda = 0;fit = [fit; sqrt(tp.sub.-- energy(xmin)) / (n.sub.-- col+1)];lambda = tmp.sub.-- lambda;endj.sub.-- plot = j.sub.-- plot + 1;figure (2)subplot(size(lambda list, 2), 1, j.sub.-- plot);plot (fit)title([`rms error, for weight: ` num2str(lambda)]);endfile: /homes/leventon/Matlab/tp.sub.-- energy.m%% tp.sub.-- energy.m%% The energy function we want to optimize%% July 16, 1997 freeman createdfunction energy = tp.sub.-- energy(x)% fprintf(1,`f`);global y M lambda alpha ncalls;e1 = sum((y - M * x). 2);e2 = lambda * sum(abs(x) . alpha);energy = e1 + e2;ncalls = ncalls + 1;if (mod(ncalls-1,100)==0)disp(sprintf(`%d %e %e %e`, ncalls, e1, e2, energy))endfile: /homes/leventon/Matlab/tp.sub.-- energy2.m%% tp.sub.-- energy.m%% The energy function we want to optimize%% July 16, 1997 freeman createdfunction energy = tp.sub.-- energy(x)% fprintf(1,`f`);global y M lambda alpha VSibeta = VSi * x;energy = sum((y - M * x). 2) + lambda * sum(abs(beta). alpha);% energy = sum(abs(beta). alpha);% energy = 0;file: /homes/leventon/Matlab/tp.sub.-- grad.m%% tp.sub.-- grad.m%% The gradient of the energy function we want to optimize%% July 16, 1997 freeman createdfunction grad = tp.sub.-- grad(x)global y M lambda alphafprintf(1, `d`);grad = 2 * M' * M * x - 2 * M' * y + lambda * alpha *(abs(x). (alpha-1)) .* sign(x);file: /homes/leventon/Matlab/trackobj.mfunction [DATA, MODELS] = trackobj(Im, 10, iM)figure(1);N = length(Im);count = 1;while (1)[m, p] = imSelect(Im{i0});if (isempty(m))break;endmodel{count} = m;range{count} = p(1,:);numModelpts(count) = length(find(m==0));count = count + 1;endnObj = count - 1;maxModels = 7;numModels = ones(nfOj, 1);saveModels = ones(nObj, 1);objectMode = ones(nObj, 1);figure(2);clfcolormap (gray (16));for i = i0:iMfor j = 1:nObjp = range{j};MODELS{i-i0+1, j} = model{j};[range{j}, v] = haus(model{j}, Im{i}, `r`, p(1)-10, p(2)-10,p(1)+10, p(2)+10);DATA(i-i0+1, j, 1) = v;% if (v < 0.4)if (v < 0.6)disp(`Lost track of an object!`);objectMode(j) = 0;endendsuperimp({model{find(objectMode)}}, Im{i},{range{find(objectMode)}}, 1);hold onplot([0.501 size(Im{i},2)], [0.501 size(Im{i},1)], `k.`);print(`-dpsc`, sprintf(`track-%02d.ps`, i-i0+1));hold offfigure (2)ifor j = 1:nObjp = range{j};if (objectMode(j))oldn = length(find(model{j}==0));if (oldn < numModelPts(j)) m = imSelect(Im{i}, [p-1; (p+size(model{j}))]); [newmodel, amt] = autocrop(and(dilate(pad(model{j}==0, 1, 1,1), ones(5)), m==0) == 0, 1);else m = imSelect(Im{i}, [p; (p+size(model{j})-1)]); [newmodel, amt] = autocrop(and(dilate(model{j}==0, ones(5))m==0) == 0, 1);endnewn = length(find(newmodel==0));DATA(i-i0+1, j, 2) = newn;if (((sum(amt) > 3) .vertline. (DATA(i-i0+1,j,1) < 0.7) .vertline.(newn/old >=0.85)) & (isempty(newmodel)==0)) subplot(nObj,maxModels, (j-1)*maxModels+numModels(j)); imagesc(model{j}); axis imageaxis off numModels(j) = numModels(j)+1; saveModels(j, numModels(j)) = i;if (numModels(j) > maxModels) numModels(j) = numModels(j) - 1; endendif (newn/numModelpts(j) >= 0.5) model{j} = newmodel;else disp(`Warning: Model cut in half: considered lost!`); objectMode(j) = 0;endsubplot(nObj ,maxModels,(j-1)*maxModels+numModels(j));imagesc(model{j});axis imageaxis offendendfigure(3);plot(DATA(:,:,1));axis([1 iM-i0+1 0 1]);figure(4);plot(DATA(:,:,2));axis([1 iM-i0+1 0 max(numModelpts)]);figure(1)pause(0.1);endfile: /homes/leventon/Matlab/train.sub.-- rbf.m%% train.sub.-- rbf.m%% Find the coefficients for a radial basis function network.%% Aug. 1, 1997 freeman created.%% Inputs:%% x.sub.-- train: a matrix of input variable training vectors.%% y.sub.-- train a matrix of the corresponding output training vectors.%% sigma: the regularization (stiffness) parameter.function [alpha, D, sq.sub.-- dist] = train.sub.-- rbf(y.sub.-- train,x.sub.-- train, sigma,sq.sub.-- dist.sub.-- orig);%% set up matrix D. Each column is the column of expontiated%% distances to the ith input training vector from each of the%% others.n.sub.-- train = size(x.sub.-- train,2);D = zeros(n.sub.-- train, n.sub.-- train);if (nargin < 4)sq.sub.-- dist.sub.-- orig = [ ];endif (isempty(sq.sub.-- dist.sub.-- orig))for j = 1:n.sub.-- trainsq.sub.-- dist(:,j) = sum((xtrain - x.sub.-- train(:,j) * ones(1,size(x.sub.-- train,2)) ). 2));D(:,j) = (exp(- sq.sub.-- dist(:,j) / (2 * sigma) 2));endelsesq.sub.-- dist = sq.sub.-- dist.sub.-- orig;for j = 1:n.sub.-- trainD(:,j) = (exp(- sq.sub.-- dist(:,j) / (2 * sigma) 2));endendBACKSOLVING = 1alpha = y.sub.-- train / D;file: /homes/leventon/Matlab/transpose2.mfunction B = transpose2(A)ndim = length(size(A));p = 1:ndim;p([1 2]) = p([2 1]);B = permute(A, p);file: /homes/leventon/Matlab/upsample.mfunction J = upsample(I, s)[m, n, p] = size(I);J = zeros(s*m, s*n, p);for i = 0:s-1J((1:s:s*m)+i, 1:s:s*n, :) = I;endfor i = 1:s-1J(1:s*m, (i:s:s*n)+i, :) = J(1:s*m, 1:s:s*n, :)endfile: /homes/leventon/Matlab/vectorize.mfunction v = vectorize(M)[nr, nc] = size(M);v = zeros(nr * nc, 1);for i = 1:nr;b = (i - 1) * nc + 1;e = i * nc;v(b:e) = M(i, :);endfile: /homes/leventon/Matlab/viewang.mfunction viewang(model, data, T, deep)if ((nargin < 2) .vertline. (isempty(data)))data = zeros(23, 1);endif (nargin < 3)T = 1:size(data, 2);endif (nargin < 4)deep = 1000000;endheld = ishold;xmin = min(data(1,:)) - 50;xmax = max(data(1,:)) + 50;ymin = min(data(2,:)) - 50;ymax = max(data(2,:)) + 50;zmin = min(data(3,:)) - 50;zmax = max(data(3,:)) + 50;xdif = xmax - xmin;ydif = ymax - ymin;zdif = zmax - zmin;[lose, which] = sort([xdif ydif zdif]);which(3) = 6;if (which(3) == 1)add = (xdif - ydif)/2;ymin = ymin - add;ymax = ymax + add;add = (xdif - zdif)/2;zmin = zmin - add;zmax = zmax + add;endif (which(3) == 2)add = (ydif - xdif)/2;xinin = xmin - add;xmax = xmax + add;add = (ydif - zdif)/2;zmin = zmin - add;zmax = zmax + add;endif (which(3) == 3)add = (zdif - xdif)/2;xmin = xmin - add;xmax = xxnax + add;add = (zdif - ydif)/2;ymin = yinin - add;ymax = ymax + add;endaxisdat = [xmin xmax ymin ymax zmin zmax];for t = Tmodel = setModelData(model, data(:,t));parts = model{8};pos = zeros(1,6);for i = 2:7if (isempty(model{i}) == 0)pos(i-1) = model{i};endendn = size(parts, 2);for i = 1:ndrawpart(parts{i}, pos(1:3), pos(4:6), deep);hold onendaxis(axisdat);view(190, 30);grid onpause(0.1);hold offendfile: /homes/leventon/Matlab/viewaoa.mfunction viewaoa(Ac, cl, frames, lednum, viewang, axisval, titleval,filespec)%%if (iscell(Ac))Al = Ac{1};elseA1 = Ac;clear AcAc{1} = A1;end[lose, ncells] = size(Ac);f = A1(1,1);m = A1(1,2);hz = A1(1,3);[n, lose] = size(A1);A1 = A1(2:n,:);xmin = min(A1(:,1));xmax = max(A1(:,1))+ 0.1;ymin = min(A1(:,2));ymax = max(A1(:,2))+ 0.1;zmin = min(A1(:,3));zmax = max(A1(:,3))+ 0.1;for ci = 1:ncellsAc{ci} = Ac{ci}(2:n,:);xmin = min(xmin, min(Ac{ci}(:,1)));xmax = max(xmax, max(Ac{ci}(:,1)));ymin = min(ymin, min(Ac{ci}(:,2)));ymax = max(ymax, max(Ac{ci}(:,2)));zmin = min(zmin, min(Ac{ci}(:,3)));zmax = max(zmax, max(Ac{ci}(:,3)));endn = n - 1;if (n .sup..about. = f * m)error(`Bad AOA`)endif (nargin < 2)cl = [ ];endif (nargin < 3)frames = 1:f;endif (nargin < 4)lednum = [ ];endif (nargin < 5)viewang [-37.5 30];endif (nargin < 6)axisval = [xmin xmax ymin ymax zmin zmax];endif (nargin < 7)titleval = ` `;endif (nargin < 8)filespec = [ ];endif (isempty(frames))frames = 1:f;endif (isempty(axisval))axisval = [xmin xmax ymin ymax zmin zmax];endif (isempty(viewang))viewang = [-37.5 30];endcolori = `brgmckybrgmckybrgmckybrgmckybrgmcky`;% colori = `gbrgmckybrgmckybrgmckybrgmckybrgmcky`;% colori = `crgmckybrgmckybrgmckybrgmckybrgmcky`);for i = framesb = (i-1)*m+1;e = i*m;hold offfor ci = 1 : ncellsA = Ac{ci};% A(:,[2 3]) = A(:,[3 2]);Plot3(axisval(1:2), axisval(3:4), axisval(5:6), `k.`);hold onplot3(axisval([1 1]), axisval(3:4), axisval(5:6), `k.`);plot3(axisval([1 1]), axisval(3:4), axisval(5:6), `k.`);plot3(axisval([1 1]), axisval(3:4), axisval(5:6), `k.`);plot3(axisval([2 2]), axisval(3:4), axisval(5:6), `k.`);plot3(axisval([2 2]), axisval(3:4), axisval(5:6), `k.`);plot3(axisval([2 2]), axisval(3:4), axisval(5:6), `k.`);axis equal;grid on% axis offplot3(A(b:e,1), A(b:e,2), A(b:e,3), [colori(ci) `o`]);title(titleval);% title(sprintf(`Frame %d`, i));for j = lednumplot3(A(j+b,1), A(j+b,2), A(j+b,3), `r*`);endn = size(cl);for j = 1:ndrawseg3(A(b+cl(j,1), :), A(b+cl(j,2), :), colori(ci));endendview(viewang(1), viewang(2));xiabel(`x`); ylabel(`y`); zlabel(`z`);if (length(fraxnes) > 1)pause(0.1);endif (isempty(filespec) == 0)fn = sprintf(filespec, i);print(`-dpsc`, fn);endendfile: /homes/leventon/Matlab/viewaoa2.mfunction viewaoa2(Ac, cl, frames, lednum, axisval, titleval, filespec)%%%% 10/10/97 wtf modified to plot orientation elements, too.%% Add two columns to the input data for the orientation of each% of the input points to be plotted. Will plot line segments oriented%%%% that way. For the array A1 below: [f, m, dontcare,% dontcare;% pix.sub.-- t1, ply.sub.-- t1, ori1x.sub.-- t1,% ori1y.sub.-- t1;% p2x.sub.-- t1, p2y.sub.-- t1, ori2x.sub.-- t1,% ori2y.sub.-- t1;% . . .% p1x.sub.-- t2, p1y.sub.-- t2, ori1x.sub.-- t2,% ori1y.sub.-- t2;% p2x.sub.-- t2, p2y.sub.-- t2, ori2x.sub.-- t2,% ori2y.sub.-- t2;% . . .% . . .% ]if (iscell(Ac))A1 = Ac{1};elseA1 = Ac;clear AcAc{1} = A1;end[lose, ncells] = size(Ac);f = A1(1,1);m = A1(1,2);[n, width] = size(A1);A1 = A1(2:n, :)xmin = min(A1(:,1));xmax = max(A1(:,1));xmin = min(A1(:,2));xmax = max(A1(:,2));for ci = 1:ncellsAc{ci} = Ac{ci}(2:n,:);xmin = min(xmin, min(Ac{ci}(:,1)));xmax = max(xmax, max(Ac{ci}(:,1)));ymin = min(ymin, min(Ac{ci}(:,2)));ymax = max(ymax, max(Ac{ci}(:,2)));endn = n - 1;if (n .sup..about. = f * m)error(`Bad AOA`)endif (nargin < 2)cl = [ ];endif (nargin < 3)frames = 1:f;endif (nargin < 4)lednum = [ ];endif (nargin < 5)axisval = [xmin xmax ymin ymax];endif (nargin < 6)titleval = ` `;endif (nargin < 7)filespec = [ ];endif (isempty(frames))frames = 1:f;endif (isempty(axisval))axisval = [xmin xmax ymin ymax];endcolori = `brgmckybrgmckybrgmckybrgmckybrgmcky`;% colori = `gbrgmckybrgmckybrgrackybrgmckybrgmcky`;% colori = `ybcrgmckybrgmckybrgmckybrgmckybrgmcky`;for i = framesb = (i-1)*m+1;e = i*m;hold offfor ci = 1 : ncellsA = Ac{ci};plot(axisval(1:2), axisval(3:4), `w.`);hold onaxis image;grid onplot(A(b:e,1), A(b:e,2), [colori(ci) `o`]);%% width == 4 means we're ploting "pulses"--orientations as well as%% points.if (width == 4)line([A(b:e, 1) + A(b:e, 3), A(b:e, 1) - A(b:e, 3)]`, . . .[A(b:e, 2) + A(b:e, 4), A(b:e, 2) - A(b:e, 4)]`);set(gca, `YDir`, `reverse`);endtitle(titleval);for j = lednumplot(A(j+b,1), A(j+b,2), [`r*`]);endn = size(cl);for j = 1:nplot([A(b+cl(j,1), 1) A(b+cl(j,2), 1)], [A(b+cl(j,1), 2)A(b+cl(j,2), 2)], colori(ci));endendxlabel(`x`); ylabel(`y`);if (length(frames) > 1)pause(0.1);endif (isempty(filespec) == 0)fn = sprintf(filespec, i);print(`-dpsc`, fn);endendfile: /homes/leventon/MatlabIviewaoa.sub.-- hacked.mfunction viewaoa(Ac, cl, frames, lednum, viewang, axisval, titleval,filespec, expand)%%if (iscell(Ac))A1 = Ac{i};elseA1 = Ac;clear AcAc{1} = A1;end[lose, ncells] = size(Ac);f = A1(1,1);m = A1(1,2);hz = A1(1,3);[n, lose] = size(A1);A1 = A1(2:n,:);xmin = inin(A1(:,1));xmax(A1(:,1))+0.1;ymin = min(A1(:,2));ymax = max(A1(:,2))+0.1;zmin = min(A1(:,3));zmax = zmax(A1(:,3))+0.1;for ci = 1:ncellsAc{ci} = Ac{ci}(2:n,:);xmin = min(xmin, min(Ac{ci}(:,1)));xmax = max(xmax, max(Ac{ci}(:,1)));ymin = min(ymin, min(Ac{ci}(:,2)));ymax = max(ymax, max(Ac{ci}(:,2)));zmin = min(zmin, min(Ac{ci}(:,3)));zmax = max(zmax, max(Ac{ci}(:,3)));endn = n - 1;if (n .sup..about. = f * m)error(`Bad AOA`)endif (nargin < 2)cl = [ ];endif (nargin < 3)frames = 1:f;endif (nargin < 4)lednum = [ ];endif (nargin < 5)viewang = [-37.5 30];endif (nargin < 6)axisval = [xmin xmax ymin ymax zmin zmax];endif (nargin < 7)titleval = ` 2 ;endif (nargin < 8)filespec = [ ];endif (nargin < 9)expand = [0 0 0];endif (isempty(frames))frames = 1:f;endif (isempty(axisval))axisval = [xmin xmax ymin ymax zmin zmax];endif (isempty(viewang))viewang = [-37.5 30];endcolori = `brgmckybrgmckybrgmckybrgmckybrgmcky`;% colori = `gbrgmckybrgmckybrgmckybrgmckybrgmcky`;% colori = `crgmckybrgmckybrgmckybrgmckybrgmcky`;colors = colormap;count = 0;for i = framescount = count + 1;factor = expand * count;b = (i-1)*m+1;e = i*m;for ci = 1 : ncellsA = Ac{ci};plot3(axisval(1:2), axisval(3:4), axisval(5:6), `k.`);hold onplot3(axisval([1 1]), axisval(3:4), axisval(5:6), `k.`);plot3(axisval([1 1]), axisval(3:4), axisval(5:6), `k.`);plot3(axisval([1 1]), axisval(3:4)) axisval(5:6), `k.`);plot3(axisval([2 2]), axisval(3:4), axisval(5:6), `k.`);plot3(axisval([2 2]), axisval(3:4), axisval(5:6), `k.`);plot3(axisval([2 2]), axisval(3:4), axisval(5:6), `k.`);axis equal;grid onH = plot3(A(b:e,1)+factor(#), A(b:e,2)+factor(2), . . .A(b:e,3)+factor(3),[colori(ci) `o`]);set(H, `Color`, colors(count,:));title(titleval);for j = lednumplot3(A(j+b,1), A(j+b,2), A(j+b,3), `r*`);endn = size(cl);for j = 1:nH = drawseg3(A(b+cl(j,1), :) + factor, . . .A(b+cl(j,2), :) + factor, colori(ci));set(H, `Color`, colors(count, :));endendview(viewang(1), viewang(2));xlabel(`x`); ylabel(`y`); zlabel(`z`);if (length(frames) > 1)pause(0.1);endif (isempty(filespec) == 0)fn = sprintf(filespec, i);print(`-dpsc`, fn);endendfile: /homes/leventon/Matlab/viewthem.mviewaoa(E1{1}, dancecl, [ ], [ ], [30 30], [ ], `E-vector 1`);pause(1)viewaoa(E1{2}, dancecl, [ ], [ ], [30 30], [ ], `E-vector 2`);pause(1)viewaoa(E1{3}, dancecl, [ ], [ ], [60 30], [ 295.6468 511.9456-59.3639 156.9350 -49.5598 66.7390], `E-vector 3`);pause(1)viewaoa(E1{4}, dancecl, [ ], [ ], [250 30], [-197.3705 250.5825-223.0159 224.9371 -63.0932 60.8598], `E-vector 4`);viewaoa(Ei{5}, dancecl, [ ], [ ], [-37 30], [ ], `E-vector 5`);pause(1)viewaoa(aoaNegate(E1{6}), dancecl, [ ], [ ], [-37 30], [ ], `E-vector6`);pause(1)viewaoa(E1{7}, dancecl, [ ], [ ], [-37 30], [ ], `E-vector 7`);pause(1)viewaoa(E1{8}, dancecl, [ ], [ ], [-37 30], [ ], `E-vector 8`);pause(1)viewaoa(E1{9}, dancecl, [ ], [ ], [-37 30], [ ], `E-vector 9`);pause(1)viewaoa(E1{10}, dancecl, [ ], [ ], [-37 30], [ ], `E-vector 10`);pause(1)file: /homes/leventon/Matlab/vmNorm.mfunction v = vmNorm(M, dim)% Take the norm of each row (colm) of M and return them in v% By default dim=1, which means the norm is taken over the rows.if (nargin < 2)dim = 1;end[m, n] = size(M);if (dim == 1)for i=1:nv(i) = norm(M(:,i));endelsefor i=1:mv(i) = norm(M(i,:));endv = v';endfile: /homes/leventon/Matlab/wave.mfunction b = wave(a, w, d)%% Weighted AVErage . . .nd = ndims(a);if ((d < 1) .vertline. (d >nd))error(`Bad Dimension`)elseif (d == 1)p = 1:nd;elseif (d == nd)p = [nd 1:nd-1];elsep = [d 1:d-1 d+1:nd];endb = permute(a, p);w = diag(diag(w));n = length(w);s = size(b);s(1) = 1;c = zeros(s);for i=1:nc(1, :, :, :, :, :, :, :, :, :) = c(1, :, :, :, :, :, :, :, :, :) +w(i) * b(i, :, :, :, :, :, :, :, :, :);endc = c / sum(w);b = ipermute(c, p);file: /homes/leventon/Matlab/wave2.mfunction b = wave(a, w)[n, m] = size(a)b = zeros(n, 1);for i = 1:ntw = 0;for j = 1:mif (isnan(a(i,j)) == 0)tw = tw + w(j);b(i) = b(i) + w(j) * a(i,j);endendif (tw == 0)b(i) = 0;elseb(i) = b(i) / tw;endendfile: /homes/leventon/Matlab/workaoa.mfunction w = workaoa(A)nfr = A(1,1);nmk = A(1,2);n = size(A,1);A = A(2:n, :);B = zeros(3, nmk, nfr);for t = 1:nfrB(L,Lt) = A((t-1)*nmk+1:t*nmk, :)`;endif (1)w = 0;for t = 2:nfr-1for i = 1:nmka = B(:,i,t+1) - 2*B(:,i,t) + B(L,i,t-1);d = (B(:,i,t+1) - B(:,i,t-1))/2;w = w + dot(a,d);endendendif (0)w = 0;for t = 2:nfr-1a = - sum(B(:,:,t+1),2) + 2*sum(B(:,:,t),2) - sum(B(:,:,t-1),2);d = (sum(B(:,:,t+1),2) - sum(B(:,:,t-1), 2))/2;w = w + dot(a,d);endw = w / nmk / nmk;endfile: /homes/leventon/Matlab/worldToCam.mfunction c = worldToCam(w, theta)if (nargin < 2)theta = 0;endif (size(w, 2) == 3)p = zeros(size(w, 1), 3);p(:,1) = 160-((w(:,1)-43.19)*(42-125)/(52.19+87.9) + 54);p(:,2) = (w(:,2)-43.19)*(42-125)/(52.19+879) + 54;p(:,3) = (w(:,3)-40)*(47-30)/(40-70) + 44;c = zeros(size(w,1), 2);c(:,2) = p(:,3);c(:,1) = p(:,1)*cos(theta) + p(:,2)*sin(theta);elseif (size(w, 2) == 2)p = zeros(size(w, 1), 3);p(:,1) = 160-((w(:,1)-43.19)*(42-125)/(52.19+87.9) + 54);p(:,3) = (w(:,2)-40)*(47-30)/(40-70) + 44;c = zeros(size(w, 1), 2);c(: ,2) = p(: ,3)c(:,1) = (p:,1);endfile: /homes/leventon/Matlab/wsum.mfunction b = wsum(a, w, d)nd = ndims(a);if ((d < 1) .vertline. (d > nd))error(`Bad Dimension`)elseif (d == 1)p = 1:nd;elseif (d == nd)p = [nd 1:nd-1];elsep = [d 1:d-1 d+1:nd];endb = permute(a, p);w = diag(diag(w));n = length(w);s = size(b);s(1) = 1;c = zeros(s);for i=1:nc(1, :, :, :, :, :, :, :, :, :) = c(1, :, :, :, :, :, :, :, :, :) +w(i) * b(i, :, :, :, :, :, :, :, :, :);b = ipermute(c, p);__________________________________________________________________________

Having above indicated a preferred embodiment of the present invention, it will occur to those skilled in the art that modifications and alternatives can be practiced within the spirit of the invention. It is accordingly intended to define the scope of the invention only as indicated in the following claims.

Claims

1. A system for reconstructing 3-dimensional motion of a human figure from a monocularly viewed image sequence of a person, comprising:

means for digitizing said monocularly viewed image sequence;

means for providing 3-d motion training data including representative motions and markers therefor as a training image sequence;

means for grouping said 3-d motion training data and said digitized image sequence into temporal segments, each temporal segment including a predetermined number of frames;

means, coupled to said means for digitizing and said means for grouping, for optimizing said temporal segments of said 3-d motion training data according to said markers so as to find a linear combination of said temporal segments of said 3-d motion training data corresponding to said monocularly viewed image sequence.

2. The system of claim 1, and further including

means for generating a 2-d stick figure corresponding to said estimated 3-d coordinates,

means for overlaying said 2-d stick figure on said digitized image sequence;

means for displaying a result thereof,

a user input device and

means, responsive to said user input device, for correcting the result, and wherein said means for optimizing includes means for correcting said estimated 3-d coordinates in accordance with said corrected result.

3. The system of claim 1, wherein said optimizing means includes means for implementing ##EQU5## where E(.alpha.) is the energy function to be minimized by the optimal coefficients .alpha. to be found, R is the vector of sensor responses over time from the image data, where the function f converts .alpha. body motion coefficients to predicted sensor responses, where I.sub.i is the ith point position specified by the user, and P.sub.i projects the .alpha. coefficients onto the corresponding ith stick figure part 2-d position, where .LAMBDA..sup.-1 is an inverse of a gaussian probability distribution of a covariance, and where .lambda..sub.1 and .lambda..sub.2 are constants which reflect the weights of the image data, the priors over human motions, and the interactively specified 2-d point matches.

4. The system of claim 3, wherein each motion segment has a number of frames and further including means for substituting ##EQU6## for (R-f(.alpha.)).sup.2, where t is an index of the frame time of each frame of the motion segment, I.sub.i is a vector of the image intensities at time t at the 2-d position of its vector argument, and M(.alpha.) is the 2-d positions of each marker for the human motion represented by the linear combination of training segments given by the combination coefficients .alpha., to minimize image intensities at the marker positions from frame to frame and to encourage said character model to track the motion of said person.

5. The system of claim 1 including means for estimating motion segments sequentially by overlapping selected motion segments to obtain multiple estimated 3-d coordinates for a given period of time, thus to provide 3-d reconstruction for a predetermined frame by computing the average of estimates over all overlapping motion segments.

6. The system of claim 1, wherein said temporal segments are divided into subsets, said optimizing means including means for selecting those subsets that are closest to the current best estimate of the segment to be estimated by said optimizing means, thereby to avoid overfitting and thus improve the estimation of said estimated 3-d coordinates.

7. The system of claim 1 further comprising:

means for outputting estimated 3-d coordinates of the corresponding markers;

means for fitting a character model to the estimated 3-d coordinates of said markers in said linear combination of said temporal segments of said 3-d motion training data; and

means for displaying said character model and motion thereof.

8. A method for estimating 3-d motion of a moving input figure in a sequence of input frames; comprising the step of:

providing a sequence of training frames, the training frames including a moving training figure, the moving training figure including markers;

grouping the training frames into a plurality of training segments, each training segment including a predetermined number of temporally adjacent frames of the sequence;

determining 3-d motion of the training figure for each training segment from the markers;

grouping the input frames into a plurality of input segments, each input segment including a predetermined number of temporally adjacent input frames;

locating a linear combination of training segments that substantially corresponds to the sequence of input segments; and

extracting the 3-d motions of the training figure in the linear combination of training segments to estimate the 3-d motion of the moving input figure.

9. The method of claim 8 further comprising:

fitting a character model to the estimated motion of the moving input figure; and

displaying the moving character model.

10. The method of claim 8 further comprising:

fitting a stick figure to the estimated motion of the moving input figure; and

correcting the estimated motion of the input figure by manually manipulating the fit of the stick figure.