Patent Grant
8363961
1. Field of the Invention
The present invention is directed generally to digital data processing, and more particularly to analysis of high-dimensionality data.
2. Description of the Related Art
Machine-based classification and identification of images based on digital image data is a complex problem. Compared to other types of digitized data, image data for any but the simplest images tends to be high in bandwidth and complexity. Techniques to identify features in images, such as the objects or surfaces represented in an image, often rely on algorithms that attempt to identify points in an image that bear some relationship to one another and to segregate these identified points from other sets of points. Such algorithms may be referred to as clustering algorithms.
In order to robustly represent the rich variety of features that may be present in image data, representations of image data may be highly dimensional. For example, image data may reflect numerous image properties beyond merely geometric location within a Cartesian coordinate space. Depending on the approach used, individual image data points may have dozens or hundreds of dimensions.
To be effective in the context of imaging, clustering algorithms thus need to readily extend to highly dimensional data. However, many existing clustering techniques are not robust, in that they are sensitive to noise in the image data and/or small variations in the parameters that govern the algorithm's performance. Moreover, complex images may include vast numbers of data points of high dimensionality, and many existing clustering techniques fail to scale well as the data set size increases.
Various embodiments of systems, methods, and computer-readable storage media for clustering highly-dimensional data are presented. In one embodiment, a computer-implemented method includes generating a mode m of a set P including a plurality of data points defined within a multidimensional space. Generating the mode may include (a) for a point p selected from within set P, identifying a set S including a plurality of points within set P that neighbor point p according to a distance metric defined on the multidimensional space; (b) generating a centroid point c of set S, wherein centroid point c is a member of set S; (c) determining whether a distance between the centroid point c and the point p satisfies a threshold value; (d) in response to determining that the distance between the centroid point c and the point p does satisfy the threshold value, returning centroid point c as a mode m of set P and terminating, and in response to determining that the distance between the centroid point c and the point p does not satisfy a threshold value, assigning centroid point c to point p and repeating (a) through (d). Each of (a) through (d) may be implemented by one or more computer systems.
While the disclosure is susceptible to various modifications and alternative forms, specific embodiments are shown by way of example in the drawings and are herein described in detail. It should be understood, however, that drawings and detailed description thereto are not intended to limit the disclosure to the particular form disclosed, but on the contrary, the disclosure is to cover all modifications, equivalents and alternatives falling within the spirit and scope of the present invention as defined by the appended claims.
In the following detailed description, numerous specific details are set forth to provide a thorough understanding of claimed subject matter. However, it will be understood by those skilled in the art that claimed subject matter may be practiced without these specific details. In other instances, methods, apparatuses or systems that would be known by one of ordinary skill have not been described in detail so as not to obscure claimed subject matter.
Some portions of the following detailed description are presented in terms of algorithms or symbolic representations of operations on binary digital signals stored within a memory of a specific apparatus or special purpose computing device or platform. In the context of this particular specification, the term specific apparatus or the like includes a general purpose computer once it is programmed to perform particular functions pursuant to instructions from program software. Algorithmic descriptions or symbolic representations are examples of techniques used by those of ordinary skill in the signal processing or related arts to convey the substance of their work to others skilled in the art. An algorithm is here, and is generally, considered to be a self-consistent sequence of operations or similar signal processing leading to a desired result. In this context, operations or processing involve physical manipulation of physical quantities. Typically, although not necessarily, such quantities may take the form of electrical or magnetic signals capable of being stored, transferred, combined, compared or otherwise manipulated. It has proven convenient at times, principally for reasons of common usage, to refer to such signals as bits, data, values, elements, symbols, characters, terms, numbers, numerals or the like. It should be understood, however, that all of these or similar terms are to be associated with appropriate physical quantities and are merely convenient labels. Unless specifically stated otherwise, as apparent from the following discussion, it is appreciated that throughout this specification discussions utilizing terms such as “processing,” “computing,” “calculating,” “determining” or the like refer to actions or processes of a specific apparatus, such as a special purpose computer or a similar special purpose electronic computing device. In the context of this specification, therefore, a special purpose computer or a similar special purpose electronic computing device is capable of manipulating or transforming signals, typically represented as physical electronic or magnetic quantities within memories, registers, or other information storage devices, transmission devices, or display devices of the special purpose computer or similar special purpose electronic computing device.
Many types of computational problems involve the processing of sets of data points that are defined within a high-dimensional space. For example, in the context of image processing, a polygonal surface of an imaged 3-dimensional object may be characterized by a set of coordinates defining the location of the surface's vertices within a 3-dimensional coordinate space. Additionally, the surface may have properties such as color, transparency, reflectivity, or other image properties defined according to the image model being employed. For a particular surface, each of these various properties may be discretely characterized by parameters within a corresponding feature space. Thus, a full description of the properties of a particular image surface may involve dozens or hundreds of independently-defined data points, each falling along an axis within a high-dimensional feature space. It is noted that, for purposes of explanation, image processing is used frequently herein as an example application. However, the techniques described herein are in no way limited to image processing, and may be used in any suitable application context on any suitable type of data set (e.g., audio data, video data, seismic data, statistical data, etc.).
Similarly, other systems of feature representation may involve feature spaces of high dimensionality. For example, the scale invariant feature transform (SIFT) algorithm may be applied to an image to generate image keypoints and associated descriptors. In one implementation, SIFT descriptors may be defined within a 128-dimensional feature space.
Given a set of data points defined within a high-dimensional feature space, one common data processing problem is the task of clustering the data points into various groups. Clustering problems may arise, for example, in the context of attempting to identify particular features or regions within an image, or attempting to classify an unknown image with respect to a dictionary or database of images. For instance, an image may depict a number of different objects that are distinct and readily identifiable by the human eye. In order to automate the recognition of depicted objects by a computer, a clustering approach may be applied to a set of data points that is representative of the entire image, with the goal of segregating the data points into individual groups. If the clustering is successful, data points within the resultant groups should bear more feature similarities with each other than with data points in other groups. Thus, the resultant groups may correspond to particular regions or objects within the original image, which may facilitate the subsequent recognition or classification of the image.
One particular class of clustering algorithm may be referred to as mode-seeking algorithms. Generally speaking, mode-seeking algorithms attempt to differentiate regions of the feature space having a comparatively high density of data points from regions of the feature space that are comparatively sparsely populated with data points. High-density regions of the feature space may also be referred to as modes. In some embodiments, mode-seeking algorithms may identify denser regions by locating maxima in the probability density field corresponding to the data points.
The mean-shift algorithm is one type of mode-seeking algorithm that may be employed in a clustering analysis. The general operation of the mean-shift algorithm on a data set is as follows: for each given point in the data set, compute the mean (i.e., the mathematical average) of all of the data points in the given point's nearest neighborhood. Then, iteratively compute the mean of all of the data points in the nearest neighborhood of the previously computed mean, such that each computed mean becomes the basis for determining the nearest neighbors used to compute the successive mean. (In other words, at each iteration, the location in the feature space around which the nearest neighbors are determined is “shifted” to the previously computed mean.) Iteration terminates when the difference in successively computed means falls below some threshold convergence value. For example, in the case of perfect convergence after a sufficient number of iterations, the location of the mean may become stationary within the feature space, such that the difference between successively computed means is zero. After the mean-shift algorithm terminates, the final location of each computed mean may correspond to a mode of the data set. Those points that converge to the same mode may be identified as belonging to the same cluster of points.
In various embodiments, the neighborhood of the given point may be defined as a region that contains a specific number of data points k that are closest to the given point (also referred to as a k-nearest neighbor problem) or a region that is defined in terms of a radius R in the feature space around the given point (also referred to as an R-nearest neighbor problem). It is noted that in the k-nearest neighbor problem, the number of nearest neighbor points is fixed at k, while the size of the region containing the nearest neighbors may vary depending on point density in the region of the given point. By contrast, in the R-nearest neighbor problem, the size of the nearest-neighbor region is fixed at R, while the number of points contained in the region may vary depending on point density in the region of the given point.
In general, the mean-shift algorithm is parameterized only in terms of the neighborhood size (e.g., k or R) and does not require prior knowledge of the number of clusters in a data set or their shape or composition. However, using the mean function as the basis for computing each iteration of the algorithm may lack robustness, in that the presence of noise (e.g., outlying points) in the data set may significantly perturb the results. As a simplistic example, consider the set of scalar data points {2, 3, 5, 5, 6, 7, 7, 7, 9}. The mean of these nine points is 5.67. If the set of data points contains a tenth outlier point having a value of 20 (e.g., as the result of sampling noise, measurement error, or another anomaly), the mean of the set shifts to 7.1—a shift of over 25% relative to the mean computed without the outlier.
Tukey Median and Locality Sensitive Hashing
With respect to scalar (i.e., 1-dimensional) data, the median is typically more robust with respect to outliers than is the mean. For example, the median of the nine-point data set in the above example is 6, whereas the median of the ten-point data set including the outlier value of 20 is 6.5 (i.e., the mean of the middle two data points in the ordered set). Thus, inclusion of the outlier value perturbs the median by less than 10%, in contrast to the 25% perturbation to the mean.
Strictly speaking, a median value is meaningful only with respect to 1-dimensional data. However, the concept of a median can be generalized to multidimensional data. Broadly speaking, a centerpoint of a multidimensional data set may be defined as a point in the multidimensional space for which any hyperplane passing through the centerpoint divides the data set into two roughly equal parts. The halfspace depth of any point in a multidimensional space refers to the minimum number of data points that lie on one side of a hyperplane through that point. The Tukey median is a centerpoint of a multidimensional data set that maximizes the halfspace depth.
Generally, the Tukey median of a set of multidimensional data points may exhibit similar robustness to the effects of outlier data, relative to the mean, as the median of a 1-dimensional data set. Thus, in some embodiments, the Tukey median of a defined neighborhood of data points may be substituted for the mean of the neighborhood of data points in the mean-shift algorithm described above. This substitution may improve the robustness of the mode-seeking algorithm with respect to anomalous data, relative to the mean-shift algorithm. Analytically determining the Tukey median of a set of data points may be computationally difficult. However, an approximation technique that is useful for identifying the nearest neighbors of a point in a multidimensional space—known as Locality Sensitive Hashing (LSH)—may be extended to generate an accurate approximation of the Tukey median of those neighboring points. Before considering the application of LSH techniques to determine the Tukey median, it is useful to consider the general features of the LSH algorithm.
At an intuitive level, the LSH algorithm operates under the premise that in general, if data points are close together within a multidimensional space, the projections of those data points onto a set of randomly selected 1-dimensional lines should also be close together. Thus, to identify the nearest neighbors of a given point in the multidimensional space, the neighbors of the given point as projected onto the 1-dimensional lines may be employed. Since identifying the projections of points onto a fixed set of lines is less computationally expensive than the brute-force technique of determining the distance from a given point to all other points within the multidimensional space, the LSH algorithm may improve the performance of locating nearest neighbors.
More formally, in one embodiment, the location of nearest neighbors using the LSH algorithm proceeds by first hashing all of the data points p in the multidimensional data set P using a set of k hash functions hi(x) (where i varies from 1 to k), each of which is divided into L bins or buckets. The choice of k and L may influence the accuracy and computational performance of the algorithm, and may be tuned for different implementations of the algorithm.
Each of the hash functions hi(x) may have the general property that if two points p and q in the multidimensional are closer together than some distance r, the probability that hi(x) will hash p and q into the same hash bucket is high. Conversely, if points p and q are separated by at least distance r, the probability that hi(x) will hash p and q into the same hash bucket is lower. More formally,
d(p,q)<r:Pr(hi(p)=hi(q))≧c1
d(p,q)≧r:Pr(hi(p)=hi(q))≧c2
where d(p, q) expresses the distance between points p and q. Under these conditions, the probability that p and q will hash to the same bucket if the distance metric is satisfied is at least c1, while the probability that these two points will hash to the same bucket if the distance metric is not satisfied is at most c2, where c1>c2.
In one embodiment, hash functions hi(x) may correspond to a set of k randomly-selected vectors or lines within the multidimensional data space, each of which is divided into L equal segments. The hash of any given point q in the data space may then be given as the dot product of q with each of the vectors corresponding to hi(x) That is, the hash of point q involves projecting point q onto each of the k vectors and determining, for each of the k vectors, which one of the L buckets of that vector the projection of q falls into.
A query to identify the nearest neighbors of point q may then occur (block 102). In response to the query, point q is hashed according to the k hash functions hi(x) to identify the buckets into which point q falls (block 104). For example, the dot product of point q with each of hi(x) may be determined, which may identify a point of projection of q along the line described by each of hi(x). For each line, this resultant point falls into segment b, a particular one of the L segments of that line. Point q may then be said to fall within bucket b of that line.
Once the buckets b of hash functions hi(x) to which point q hashes have been determined, the other points in data set P that hash to the same bucket(s) as point q may be identified as candidate neighbors (block 106). For example, the previously-generated hash data structure indicates which points of P fall within which buckets. Therefore, this data structure may be used to determine the other points of P that are members of the same bucket as q. In some embodiments, any point of P that hashes to the same bucket as q for any one of hash functions hi(x) may be considered a candidate neighbor of q. In other embodiments, to qualify as a candidate neighbor, a point of P must hash to the same bucket as q with respect to at least some number n of the hash functions hi(x), where n>1.
Once the candidate neighbors have been identified, the distance between each of the candidate neighbors and point q may be determined (block 108). Those candidate neighbors that fall within distance r of point q may be identified as the R-nearest neighbors of q (block 110).
It is noted that while computing the distance between two points in a multidimensional space may be computationally expensive, the hashing process may reduce the number of candidate neighbors for which the distance must be computed.
That is, in the brute-force approach, every point in P is a candidate neighbor of q for which a distance must be determined. By contrast, in the LSH approach, the number of candidate neighbors may be less than all points in P—in some cases, substantially less, depending on how the points are distributed within the multidimensional space and the behavior of the hash functions used. Thus, use of the LSH algorithm may scale to accommodate much larger data sets than other algorithms.
In the foregoing discussion of the LSH algorithm, it was seen that by projecting points in a multidimensional space onto a set of vectors defined within a simpler, 1-dimensional space, the problem of determining nearest neighbors may be simplified. It has been shown that the Tukey median of a set of points in a multidimensional space may also be approximated by projecting the points onto a random set of independent 1-dimensional vectors and analyzing the projections to identify the Tukey median. Thus, the problems of generating nearest neighbors and of identifying the Tukey median of a set of points overlap to the extent that 1-dimensional projective techniques may be used to address both problems.
The Tukey median of a set of points P in a multidimensional space may also be referred to as the centroid of P. In one embodiment, shown in
Centroid(P)=minƎcεP{sum∀νεV((c∘v)−median(P∘v))}
where ∘ denotes a suitable operator for projection of a point or a set of points onto a vector, such as the dot product operator.
It is noted that in this embodiment, the centroid of P is defined as a member of the set of points P. That is, the centroid of a data set has the property that it is one of the points that actually exists within the data set. By contrast, other metrics—such as the mean of a data set—may produce a point within the multidimensional space that does not correspond to any point within the data set. Defining the centroid to be a member of the data set may decrease the computational complexity and noise susceptibility of mode-seeking algorithms, relative to metrics such as the mean.
In some embodiments, the set of randomly selected 1-dimensional vectors that is used to determine the centroid may be identical to the set of hash function vectors hi(x) vectors employed by the LSH algorithm to determine nearest neighbors. In such an embodiment, the hash data structure used to store the projective hash may also be employed to retrieve the projection data, discussed above, that is used to determine the centroid. In some embodiments, the hash data structure may be augmented to represent projection data for each of vectors hi(x) with a greater precision than provided by the L hash buckets discussed above. For example, for each vector, the hash data structure may indicate data point membership in each of L hash buckets for the purpose of nearest-neighbor identification, but may also indicate data point membership in each of M buckets (M>L) for the purpose of centroid identification.
It is contemplated that in other embodiments, the sets of vectors and data structures used to determine and store centroid information may be entirely distinct from those vectors and data structures used to identify nearest neighbors.
Centroid-Shift Clustering
The concepts of using LSH-based nearest neighbor techniques and of the projection-approximated Tukey-median-based centroid discussed above may be combined and extended to yield a new mode-seeking clustering algorithm, referred to herein as centroid-shift clustering. Generally speaking, centroid-shift clustering proceeds in an iterative manner similar to the mean-shift algorithm. However, instead of determining the mean of a set of nearest neighbor points at each iteration, the centroid-shift algorithm determines the centroid. As noted above, using the centroid may yield an algorithm that is more robust to anomalous data than the mean-shift algorithm. Also, the use of hash-based nearest neighbor location algorithms such as LSH as part of the centroid-shift algorithm may enable the algorithm to accommodate larger data sets with better computational performance than other algorithms.
One embodiment of a method of centroid-shift clustering is illustrated in
In one embodiment, the general approach of the LSH algorithm discussed above with respect to
It is noted that in one embodiment of the centroid-shift algorithm, it may be unnecessary to compute the distance from p to each candidate neighbor and then eliminate those candidates that do not satisfy the distance parameter r. As noted above, the hash functions hi(x) employed in the LSH algorithm are characterized in terms of the probability that they will hash points that are within distance r from each other into the same buckets. Hash functions hi(x) may be selected such that on average, a significant majority (e.g., 80-90% in some embodiments) of the points returned by the LSH algorithm as candidate neighbors of a point q do in fact fall within distance r from q. It is true that the set of candidate neighbors may include outliers that fail to fall within distance r from q. However, as noted above, the centroid algorithm may be robust with respect to outliers, such that their inclusion in a data set may only minimally affect the resultant centroid, and the benefit of removing the outliers is outweighed by the cost of doing so. Therefore, in some embodiments, the pruning of candidate neighbors according to distance r, as shown in blocks 108-110 of
Given the set S of p and its neighbors, the centroid c of set S is determined (block 304). For example, the method of
The distance between point p and centroid c is then determined (block 306). If this distance satisfies a threshold value, then operation of the centroid-shift algorithm may be complete and point p may correspond to a mode of data set P (blocks 308-310). In one embodiment, the threshold value may be a numerical value such as zero or a nonzero value, and determining whether the distance satisfies the threshold value may include determining whether the distance is less than (or less than or equal to) the threshold value. For example, upon convergence, the centroid may become stationary, such that successively computed centroids are identical to or within a threshold value of previously computed centroids. If the distance between point p and centroid c is greater than the threshold value, then centroid c is assigned to point p (block 312) and iteration may proceed from block 302.
In one embodiment, the centroid-shift mode-seeking algorithm may be repeatedly applied using each point p within data set P (or, in some cases, a subset of P such as a random sampling) as the initial point. Once the mode has been identified for each of multiple points p, those points that converge to a common mode (either exactly, or in some cases within a threshold of error) may be identified as members of the same cluster of points. That is, the mode returned by the centroid-shift algorithm may be used as a criterion for cluster membership.
The centroid-shift algorithm has been described above in the context of points and sets of points. That is, the algorithm may determine a particular point c to be the centroid of a set of points that neighbor point p, and may iterate on the basis of shifting point p to the previously computed centroid point c. However, the centroid-shift algorithm may also be generalized to operate on clusters of points and sets of such clusters. That is, the algorithm may first identify a cluster of points P within a set of clusters S, and may then identify the set of neighboring clusters N that satisfy a distance metric r with respect to cluster P. Then, the centroid cluster C of N may generated by identifying the cluster that minimizes the error in the projective median of all clusters in N onto a set of vectors V, in a manner analogous to the technique described above with respect to locating the centroid point c of a set of points P. Iteration may then continue using the newly determined centroid cluster C.
Exemplary Computer System Embodiment
Embodiments of the systems and methods described herein may be used to analyze, characterize, and represent, digital images, e.g., digital photographs, digital art, scans of artwork, etc., and the representations of the images may be useful in image search applications.
In one embodiment, a specialized graphics card or other graphics component 456 may be coupled to the processor(s) 410. The graphics component 456 may include a graphics processing unit (GPU) 170, which in some embodiments may be used to perform at least a portion of the techniques described above. Additionally, the computer system 400 may include one or more imaging devices 452. The one or more imaging devices 452 may include various types of raster-based imaging devices such as monitors and printers. In one embodiment, one or more display devices 452 may be coupled to the graphics component 456 for display of data provided by the graphics component 456.
In one embodiment, program instructions 440 that may be executable by the processor(s) 410 to implement aspects of the techniques described herein (including any aspect of the techniques illustrated in
The memory 420 may be implemented using any appropriate medium such as any of various types of ROM or RAM (e.g., DRAM, SDRAM, RDRAM, SRAM, etc.), or combinations thereof. The program instructions may also be stored on a storage device 460 accessible from the processor(s) 410. Any of a variety of storage devices 460 may be used to store the program instructions 440 in different embodiments, including any desired type of persistent and/or volatile storage devices, such as individual disks, disk arrays, optical devices (e.g., CD-ROMs, CD-RW drives, DVD-ROMs, DVD-RW drives), flash memory devices, various types of RAM, holographic storage, etc. The storage 460 may be coupled to the processor(s) 410 through one or more storage or I/O interfaces. In some embodiments, the program instructions 440 may be provided to the computer system 400 via any suitable computer-readable storage medium including the memory 420 and storage devices 460 described above.
The computer system 400 may also include one or more additional I/O interfaces, such as interfaces for one or more user input devices 450. In addition, the computer system 400 may include one or more network interfaces 454 providing access to a network. It should be noted that one or more components of the computer system 400 may be located remotely and accessed via the network. The program instructions may be implemented in various embodiments using any desired programming language, scripting language, or combination of programming languages and/or scripting languages, e.g., C, C++, C#, Java™, Perl, etc. The computer system 400 may also include numerous elements not shown in
Image analysis module 500 may be implemented as or in a stand-alone application or as a module of or plug-in for an image processing application. Examples of types of applications in which embodiments of module 500 may be implemented may include, but are not limited to, image (including video) analysis, characterization, search, processing, and/or presentation applications, as well as applications in security or defense, educational, scientific, medical, publishing, digital photography, digital films, games, animation, marketing, and/or other applications in which digital image analysis, characterization, representation, or presentation may be performed. Specific examples of applications in which embodiments may be implemented include, but are not limited to, Adobe® Photoshop® and Adobe® Illustrator®. Module 500 may also be used to display, manipulate, modify, classify, and/or store images, for example to a memory medium such as a storage device or storage medium.
To facilitate presentation, many of the examples provided have been discussed in the context of image processing. However, it is noted that the clustering techniques described herein need not be limited to image processing applications. Rather, it is contemplated that these clustering techniques may be applied to any type of multidimensional data, regardless of what underlying phenomena the data represents. For example and without limitation, these clustering techniques may be employed with respect to audio data, electromagnetic field data, statistical data, or any other type of multidimensional data.
It is noted that the data set on which the above-described clustering techniques may be employed may be representative of a physical object or a physical phenomenon. The data set may be generated in any fashion suitable to the type of data, such as digitization of sensor data where the sensor is configured to detect visible or near-visible portions of the electromagnetic spectrum (e.g., visible light), invisible portions of the electromagnetic spectrum (e.g., radio frequencies), acoustic energy, temperature, motion, or any other physical input or combination of inputs. The data set may pass through other types of analysis and/or transformation prior to and/or subsequent to performance of clustering.
It is noted that the order in which the various operations detailed above have been described and illustrated has been chosen only to facilitate exposition. Additionally, the use of distinguishing reference numerals and reference characters such as (a), (b), (c), (d), etc., is intended only to facilitate the description of the items denoted with such reference numerals or reference characters, and does not dictate or imply any necessary ordering of the items so denoted. In various embodiments, the operations described herein may be performed in any suitable order, including orderings that differ from those described above. For example, where multiple instances of a method are to be performed on multiple different data points, the various operations that result may be serialized, parallelized, or interleaved in any suitable fashion. Further, it is contemplated that in some embodiments, various ones of the operations described above may be omitted or combined with other operations, or other additional operations may be performed.
Although the embodiments above have been described in detail, numerous variations and modifications will become apparent to those skilled in the art once the above disclosure is fully appreciated. It is intended that the following claims be interpreted to embrace all such variations and modifications.
This application claims the benefit of priority of U.S. Provisional Patent Application No. 61/105,276, filed Oct. 14, 2008, which is hereby incorporated by reference in its entirety.
| Number | Name | Date | Kind |
|---|---|---|---|
| 6584220 | Lantrip et al. | Jun 2003 | B2 |
| Entry |
|---|
| Slaney et al., “Locality-Sensitive Hasing for Finding Nearest Neighbors”, Mar. 2008, IEEE Signal Processing Magazine, pp. 128-131. |
| Ozertem et al. “Mean Shift Spectral Clustering”, Sep. 25, 2007, Elsevier, The Journal of the Pattern Recognition Society, pp. 1-15. |
| Indyk, “Locality-sensitive hashing using stable distributions” (with A. Andoni and M. Datar and N. Immorlica and V. Mirrokni), in Nearest Neighbor Methods in Learning and Vision: Theory and Practice (2006). |
| L. Zelnik-manor and P. Perona. “Self-tuning spectral clustering”. In Advances in Neural Information Processing Systems 17, pp. 1601-1608. MIT Press, 2005. |
| Andoni, Indyk, “Near-Optimal Hashing Algorithms for Approximate Nearest Neighbor in High Dimensions,” 2008. |
| E. K. Yaser Sheikh and T. Kanade. “Mode-seeking by medoidshifts”. In ICCV'07: Proceedings of the 11th International Conference on Computer Vision, pp. 1-8, Washington, DC, USA, Oct. 2007. IEEE Computer Society. |
| D. Comaniciu and P. Meer. “Mean shift analysis and applications,” In ICCV '99: Proceedings of the International Conference on Computer Vision-vol. 2, p. 1197, Washington, DC, USA, 1999. IEEE Computer Society. |
| A. Vedaldi and S. Soatto. “Quick shift and kernel methods for mode seeking”. In Proceedings of the European Conference on Computer Vision (ECCV), 2008. |
| M. Datar, N. Immorlica, P. Indyk, and V. S. Mirrokni, “Locality-sensitive hashing scheme based on p-stable distributions,” In SCG '04: Proceedings of the twentieth annual symposium on Computational geometry, pp. 253-262, New York, NY, USA, 2004. ACM. |
| B. Georgescu, I. Shimshoni, and P. Meer. “Mean shift based clustering in high dimensions: a texture classification example,” In Computer Vision, 2003. Proceedings. Ninth IEEE International Conference on, pp. 456-463 vol. 1, 2003. |
| D. M. Mount and S. Arya. “Ann: A library for approximate nearest neighbor searching”. 1997. |
| S. Paris and F. Durand. “A topological approach to hierarchical segmentation using mean shift”. Computer Vision and Pattern Recognition, 2007. CVPR '07. IEEE Conference on, pp. 1-8, Jun. 2007. |
| R. Subbarao and P. Meer. “Nonlinear mean shift for clustering over analytic manifolds”. In CVPR '06: Proceedings of the 2006 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pp. 1168-1175, Washington, DC, USA, 2006. IEEE Computer Society. |
| Shakhnarovich, Indyk, “Introduction,” in Nearest Neighbor Methods in Learning and Vision: Theory and Practice (2006). |
| Indyk,“Near-Optimal Hashing Algorithms for Approximate Near(est) Neighbor Problem,” 2006. |
| J. A. Cuesta-Albertos and A. Nieto-Reyes. The random tukey depth. Comput. Stat. Data Anal., 52(11):4979-4988, 2008. |
| Number | Date | Country | |
|---|---|---|---|
| 61105276 | Oct 2008 | US |
