Patent Application
20170013279
A video encoder compresses video information so that more information can be sent over a given bandwidth. The compressed signal may then be transmitted to a receiver having a decoder that decodes or decompresses the signal prior to display.
High Efficient Video Coding (HEVC) is the latest video compression standard, which is being developed by the Joint Collaborative Team on Video Coding (JCT-VC) formed by ISO/IEC Moving Picture Experts Group (MPEG) and ITU-T Video Coding Experts Group (VCEG). HEVC is being developed in response to the previous H.264/AVC (Advanced Video Coding) standard not providing enough compression for evolving higher resolution video applications. Similar to previous video coding standards, HEVC includes basic functional modules such as intra/inter prediction, transform, quantization, in-loop filtering, and entropy coding.
The ongoing HEVC standard may attempt to improve on limitations of the H.264/AVC standard such as limited choices for allowed prediction partitions and coding partitions, limited allowed multiple references and prediction generation, limited transform block sizes and actual transforms, limited mechanisms for reducing coding artifacts, and inefficient entropy encoding techniques. However, the ongoing HEVC standard may use iterative approaches to solving such problems.
For instance, with ever increasing resolution of video to be compressed and expectation of high video quality, the corresponding bitrate/bandwidth required for coding using existing video coding standards such as H.264 or even evolving standards such as H.265/HEVC, is relatively high. The aforementioned standards use expanded forms of traditional approaches to implicitly address the insufficient compression/quality problem, but often the results are limited.
The present description, developed within the context of a Next Generation Video (NGV) codec project, addresses the general problem of designing an advanced video codec that maximizes the achievable compression efficiency while remaining sufficiently practical for implementation on devices. For instance, with ever increasing resolution of video and expectation of high video quality due to availability of good displays, the corresponding bitrate/bandwidth required using existing video coding standards such as earlier MPEG standards and even the more recent H.264/AVC standard, is relatively high. H.264/AVC was not perceived to be providing high enough compression for evolving higher resolution video applications.
The material described herein is illustrated by way of example and not by way of limitation in the accompanying figures. For simplicity and clarity of illustration, elements illustrated in the figures are not necessarily drawn to scale. For example, the dimensions of some elements may be exaggerated relative to other elements for clarity. Further, where considered appropriate, reference labels have been repeated among the figures to indicate corresponding or analogous elements. In the figures:
One or more implementations are now described with reference to the enclosed figures. While specific configurations and arrangements are discussed, it should be understood that this is done for illustrative purposes only. Persons skilled in the relevant art will recognize that other configurations and arrangements may be employed without departing from the spirit and scope of the description. It will be apparent to those skilled in the relevant art that techniques and/or arrangements described herein may also be employed in a variety of other systems and applications other than what is described herein.
While the following description sets forth various implementations that may be manifested in architectures such as system-on-a-chip (SoC) architectures for example, implementation of the techniques and/or arrangements described herein are not restricted to particular architectures and/or computing systems and may be implemented by any architecture and/or computing system for similar purposes. For instance, various architectures employing, for example, multiple integrated circuit (IC) chips and/or packages, and/or various computing devices and/or consumer electronic (CE) devices such as set top boxes, smart phones, etc., may implement the techniques and/or arrangements described herein. Further, while the following description may set forth numerous specific details such as logic implementations, types and interrelationships of system components, logic partitioning/integration choices, etc., claimed subject matter may be practiced without such specific details. In other instances, some material such as, for example, control structures and full software instruction sequences, may not be shown in detail in order not to obscure the material disclosed herein.
The material disclosed herein may be implemented in hardware, firmware, software, or any combination thereof. The material disclosed herein may also be implemented as instructions stored on a machine-readable medium, which may be read and executed by one or more processors. A machine-readable medium may include any medium and/or mechanism for storing or transmitting information in a form readable by a machine (e.g., a computing device). For example, a machine-readable medium may include read only memory (ROM); random access memory (RAM); magnetic disk storage media; optical storage media; flash memory devices; electrical, optical, acoustical or other forms of propagated signals (e.g., carrier waves, infrared signals, digital signals, etc.); and others.
References in the specification to “one implementation”, “an implementation”, “an example implementation”, etc., indicate that the implementation described may include a particular feature, structure, or characteristic, but every embodiment may not necessarily include the particular feature, structure, or characteristic. Moreover, such phrases are not necessarily referring to the same implementation. Further, when a particular feature, structure, or characteristic is described in connection with an embodiment, it is submitted that it is within the knowledge of one skilled in the art to effect such feature, structure, or characteristic in connection with other implementations whether or not explicitly described herein.
Systems, apparatus, articles, and methods are described below related to dominant motion compensated prediction for next generation video coding.
As discussed above, the H.264/AVC coding standard while it represents improvement over past MPEG standards, it is still very limiting in choices of prediction due to the following reasons: the choices for allowed prediction partitions are very limited, the accuracy of prediction for prediction partitions is limited, and the allowed multiple reference predictions are very limited as they are discrete based on past decoded frames rather than accumulation of resolution over many frames. The aforementioned limitations of the state of the art standards such as H.264 are recognized by the ongoing work in HEVC that uses an iterative approach to fixing these limitations.
Further, the problem of improved prediction is currently being solved in an ad hoc manner by using decoded multiple references in the past and/or future for motion compensated prediction in inter-frame coding of video. This is done with the hope that in the past or future frames, there might be some more similar areas to the area of current frame being predicted than in the past frame (for P-pictures/slices), or in the past and future frames (for B-pictures/slices).
As will be described in greater detail below, some forms of prediction, such as the dominant motion compensation prediction procedures of this disclosure, may not be supportable by existing standards. The present disclosure was developed within the context of Next Generation Video (NGV) codec project to addresses the problem of designing a new video coding scheme that maximizes compression efficiency while remaining practical for implementation on devices. Specifically, a new type of prediction is disclosed herein called locally adaptive dominant motion compensated prediction (or simply dominant motion compensation prediction) that provides improved prediction which in turn reduces the prediction error, thereby improving the overall video coding efficiency.
More specifically, techniques described herein may differ from standards based approaches as it naturally incorporates significant content based adaptivity in video coding process to achieve higher compression. By comparison, standards based video coding approaches typically tend to squeeze higher gains by adaptations and fine tuning of legacy approaches. For instance, all standards based approaches heavily rely on adapting and further tweaking of motion compensated interframe coding as the primary means to reduce prediction differences to achieve gains. On the other hand, some video coding implementations disclosed herein, in addition to exploiting interframe differences due to motion, also exploits other types of interframe differences (gain, blur, registration, dominant/global motion) that naturally exist in typical video scenes, as well as prediction benefits of frames synthesized from past decoded frames only or a combination of past and future decoded frames. In NGV coding, the morphed frames used for prediction include dominant motion compensated (DMC) prediction. In some video coding implementations disclosed herein, the synthesized frames used for prediction include Super Resolution (SR) frames and PI (Projected Interpolation) frames. Besides the issue of exploiting other sources of interframe differences besides motion, some video coding implementations disclosed herein differ from standards in other ways as well.
With regard to dominant motion compensation (DMC), improving motion compensation for prediction is one of the keys to achieving higher coding efficiency in recent video coding standard and solutions. For example, with block-based motion compensation, a block (such as a 16×16 block of pixels) of a current frame being analyzed is matched during motion estimation to a similar block in a previously decoded reference frame. The shift from one frame to the other in the x and y directions with respect to a block grid, is referred to as a motion vector with ‘x’ and ‘y’ components referred to as mvx, and mvy. The motion estimation process thus involves estimating motion of blocks to determine my (mvx, mvy) for each block. The computed motion estimates are then efficiently encoded (by first differencing them with prediction, and entropy coding of the difference) and sent via bitstream to the decoder where they are decoded (by entropy decoding and adding the prediction back in) and used for motion compensation. In highly efficient compression schemes, motion estimation/compensation is performed with high accuracy (such as ¼ pixel or ⅛ pixel accuracy rather than integer pixel accuracy) by use of a fixed or adaptive interpolation filter for generation of a prediction block. Further, generally the block sizes themselves may be square or non-square (e.g. 16×8, 8×16) and of multiple sizes (e.g. 4×4, 8×8, 16×16, and others).
While H.264 includes several good ways of generating predictions derived from block-based motion vectors, this may result in two limitations: (1) block-based motion vectors, regardless of block sizes or references, are all modeled based on an assumption of translatory motion, which may disregard alternative types of motion between frames and resulting in large prediction error, and (2) block-based motion vectors, while they provide good compensation of local motion, the inherent substantial bit-cost associated with the block-based motion can limit the potential gain that may otherwise be possible. Improvements have involved use of variable size blocks which helps in reducing overhead, but the overhead reduction is still quite limited.
For video content undergoing global motion such as camera pan (translation), zoom, rotation, or in video content that has special effects (such as shearing), the block-based translator motion representation and coding of motion vectors can be particularly inefficient. Since it was realized that global motion in a video can present a challenge to block-based prediction due to a large prediction resulting from a translatory motion model, and significant amount of motion vector overhead, an alternative approach was investigated that directly estimates/compensates global motion due to its potential to adapt to nontranslatory/complex motion, and a more compact representation of motion parameters as they are needed only once per picture (note that herein the term frame is used interchangeably with the term picture). Among the choice of motion models for global motion, the two models that offer significant improvements are the affine model, and the perspective model. The affine model uses six parameters and is able to address a large range of complex motions (translation, zoom, shearing, and rotation). In the typical process, the model results in a warped frame used to form the predictions by reading the blocks on the warped frame. The perspective model is more complex than the affine model and in addition to the motions listed for affine, this method can also handle a change in perspective. Due to the higher complexity of the perspective model, it is not discussed here in detail, but in general it is applicable in the same manner as the affine model. The details for global motion compensation, at least as used by the system and process herein, are discussed below.
While use of an affine model-based global motion estimation/compensation (GME/C) was a notable improvement for scenes with global motion over use of block based translatory motion, video scenes in general can be classified into one of the three cases: (1) scenes with purely global motion, (2) scenes with purely local motion, and (3) scenes containing both local and global motion. Thus, in general, both the global and the local motion techniques needed to be combined for achieving good efficiency in video coding. MPEG-4, part2 supports a very basic combination of global and local motion techniques. Specifically it supports 16×16 luma block (and its optional 8×8 sub-block) based local motion estimation/compensation, picture based affine model developing a global motion trajectory (gmt) parameters-based motion compensation, and a 16×16 block by block-based flag for local or global motion (1 gm) that allows a choice of when to use which method.
While the MPEG-4 part 2 standard represents an improvement (due to inclusion of Global Motion and other aspects) over past MPEG or ITU-T standards, it still offers only a limited improvement in motion compensated prediction due to the following reasons. Even though a combination of local and global motion compensation is allowed, the local motion compensation occurs at a very small block size (16×16 size at most). Thus, there is considerable overhead in signaling, at a 16×16 basis, when to use local versus global motion compensation. This overhead cuts into the possible gains due to GMC. Also, since P-pictures only use one reference frame and B-pictures only use two, GMC is limited to being applied only on immediately past decoded frames. Further, global motion parameters are computed only once for the entire picture (including blocks where local motion is found to be more suitable) causing global motion parameters to be inaccurate often, especially in the case of a frame that contains a mixture of both local and global motion. Otherwise, other than using or not using global motion compensated prediction, no adjustment or correction of GMC generated prediction is possible. Lastly, the process for generating interpolation (e.g. ¼ or ⅛ pel precision) is simplistic and results in blurry prediction.
These difficulties are addressed by the new and innovative approaches used by a NGV video coding system including improved prediction by dominant motion compensation described herein. By one example, since global motion in video can present a challenge to block-based prediction (due to larger prediction resulting from a translatory motion model, and a significant amount of motion vector overhead), an alternative approach was developed that directly estimates and compensates global motion due to its potential of being able to better adapt to nontranslatory or complex motion, and a more compact representation of motion parameters is now available as needed such as once per frame. Among the choice of motion models for global motion, the two models that offer significant benefits are still the affine model, and the perspective model. The affine model uses six parameters, and is able to address a large range of complex motions, while the perspective model is more complex and flexible, but can use up to eight parameters. The affine model may be sufficient for many cases and can allow global compensation for motion of types such as translation, zoom, shear, and rotation.
While use of the affine model based global motion estimation or compensation (GME/C) was a notable improvement for scenes with global motion over use of block based translatory motion, in reality both block-based local and global motion is combined here for best coding efficiency results. Further, the affine model can also be applied for motion compensation of non-overlapping tiles, or regions/objects in a scene. This results in multiple global motion parameter sets, and the process is referred to as performing dominant motion compensation (DMC).
As used herein, the term “coder” may refer to an encoder and/or a decoder. Similarly, as used herein, the term “coding” may refer to performing video encoding via an encoder and/or performing video decoding via a decoder. For example, a video encoder and video decoder may both be examples of coders capable of coding video data. In addition, as used herein, the term “codec” may refer to any process, program or set of operations, such as, for example, any combination of software, firmware, and/or hardware that may implement an encoder and/or a decoder. Further, as used herein, the phrase “video data” may refer to any type of data associated with video coding such as, for example, video frames, image data, encoded bit stream data, or the like.
Referring to
As shown, video frames and/or other data may be transmitted from adaptive picture organizer module 104 to prediction partitions generator module 105. In some examples, prediction partitions generator module 105 first may divide a frame or picture into tiles or super-fragments or the like (herein the terms frame, picture, and image may be used interchangeably except as otherwise noted and except that a frame is used to generally refer to a frame that is not necessarily assigned a specific picture type (I, P, F, or B-pictures for example)). In some examples, an additional module (for example, between modules 104 and 105) may be provided for dividing a frame into tiles or super-fragments or the like. By one example for NGV coding, a frame may be divided into tiles of 32×32 or 64×64 pixels where 64×64 is used for all standard definition and higher resolution video for coding of all picture types (I-, P-, or F-). For low resolution sequences, 64×64 is still used for coding of I- and F-pictures, while 32×32 is used for P-pictures.
By one example, prediction partitions generator module (which also may be referred to as Pred KdTree/BiTree Partitions Generator) 105 may then divide each tile or super-fragment into potential prediction partitionings or partitions. In some examples, the potential prediction partitionings may be determined using a partitioning technique such as, for example, a k-d tree partitioning technique, a bi-tree partitioning technique, or the like, which may be determined based on the picture type (for example, I-, P-, or F-picture) of individual video frames, a characteristic of the frame portion being partitioned, or the like. By one example, if an I-picture is being coded, every tile, or almost all tiles, are further divided into KdTree based partitions that can divide a space until a set minimum size is reached, and in one dimension at a time. The options for dividing the space may include no further division, division into two equal halves, division into two parts that are ¼ and ¾ of the space, or division into two parts that are ¾ and ¼ of the space. So, with I-pictures using 64×64 as the largest size (and allowing a minimum size of 4×4), a very large number of partitions of a tile can be generated if no other constraints are imposed. For example, one constraint is to set that the first pair of cuts are pre-decided for a 64×64 tile to halve the space in both the horizontal and vertical dimension so that four 32×32 sub-tiles are formed, and then sub-partitioning each 32×32 sub-tile by KdTree partitioning. Other restrictions are also possible to reduce the number of possible partition combinations.
These partitions of an I-picture tile are referred to as prediction partitions, as each tile partition may be used for spatial prediction (directional angular prediction or other types of prediction) and coding of prediction differences. Likewise, P-picture tiles can also be partitioned in this manner for prediction except that for lower resolutions, P-picture partitions start with a 32×32 tile, and KdTree based partitions are not used, but rather a simpler Bi-Tree partitioning is used. Bi-Tree partitioning divides a space into two equal parts, one dimension at a time, alternating between the two dimensions. Further P-picture partitions are mainly predicted using motion (with one or more references) rather than spatial prediction, although some sub-partitions can use intra spatial prediction to deal with, for instance, uncovered background. For standard definition to higher resolution picture sizes, P-pictures start with 64×64 tiles before being divided. Finally, F-pictures also use Bi-Tree partitioning and start with 64×64 tiles for generating prediction partitions that mainly use motion (with one or more partitions), although some sub-partitions can also use spatial prediction (for intra coding).
Alternatively, where the methods described herein may be performed on HEVC where a largest coding unit (LCU), also called a coding tree unit (CTU), may be divided into coding tree blocks (CTBs) which are themselves divided into coding units (CUs). Such an LCU may be 64×64 pixels. Thus, a tile as used herein covers HEVC and generally refers to a large block such as an LCU or at least a block larger than a macroblock (MB) of 16×16 pixels unless the context suggests otherwise.
In NGV coding, there is much more to generation of inter prediction data than simply using motion vectors to generate prediction, and is discussed elsewhere. In P- and F-picture coding, each sub-partition's prediction is identified by including a prediction mode. The prediction modes include skip, auto, intra, inter, multi, and split. Skip mode is used to skip prediction coding when, for example, there is no, or relatively little change, from a reference frame to a current frame being reconstructed so that the pixel data need not be encoded and merely copied from one frame to the other when decoded. Auto mode is used when only partial data is needed so that for example, motion vectors may not be needed but transform coefficients are still used to code the data. Intra mode means that the frame or partition is spatially coded. Split means a frame or partition needs to be split into smaller parts or partitions before being coded. Inter mode means that multiple reference frames are determined for a current frame, and motion estimations are obtained by using each reference separately, and then the best result is used for the motion prediction data. Multi mode also uses multiple reference frames, but in this case, the motion estimation data from the multiple reference frames is combined, such as averaged, or weighted averaged, to obtain a single result to be used for the prediction.
One of the outputs of prediction partitions generator module 105 may be hundreds of potential partitionings (and more or less depending on the limits placed on the partitioning) of a tile. These partitionings are indexed as I . . . m and are provided to the encode controller 103 to select the best possible prediction partitioning for use. As mentioned, the determined potential prediction partitionings may be partitions for prediction (for example, inter- or intra-prediction) and may be described as prediction partitions or prediction blocks or the like.
In some examples, a selected prediction partitioning (for example, prediction partitions) may be determined from the potential prediction partitionings. For example, the selected prediction partitioning may be based on determining, for each potential prediction partitioning, predictions using characteristics and motion based multi-reference predictions or intra-predictions, and determining prediction parameters. For each potential prediction partitioning, a potential prediction error may be determined by differencing original pixels with prediction pixels, and the selected prediction partitioning may be the potential prediction partitioning with the minimum prediction error. In other examples, the selected prediction partitioning may be determined based on a rate distortion optimization including a weighted scoring based on number of bits for coding the partitioning and a prediction error associated with the prediction partitioning.
As shown, the original pixels of the selected prediction partitioning (for example, prediction partitions of a current frame) may be differenced with predicted partitions (for example, a prediction of the prediction partition of the current frame based on a reference frame or frames and other predictive data such as inter- or intra-prediction data) at differencer 106. The determination of the predicted partitions will be described further below and may include a decode loop 135 as shown in
Otherwise, any residuals or residual data (for example, partition prediction error data) from the differencing that indicate that the partition cannot be compensated by prediction alone (such as motion compensation alone) may be transmitted to coding partitions generator module (or by one example, coding bitree partitions generator) 107 to be further sub-partitioned into smaller partitions for transform coding (coding partitions), and particularly for P-pictures and F-pictures by one example. In P- or F-pictures or frames, in some cases where very simple content and/or large quantizer step sizes exist, the coding partitions may equal the size of the entire tile, or the coding partitions and prediction partitions may have the same size in these cases. Thus, some P- and F-picture tiles may contain no coding partitioning, one coding partitioning, or multiple coding partitionings. These coding partitions are indexed as I . . . n, and are provided to encode controller 103 to select the best possible combination of prediction and coding partitioning from the given choices.
Also, in some of these examples, such as for intra-prediction of prediction partitions in any picture type (I-, F- or P-pictures), or otherwise where prediction partitions are not further divided into coding partitions (where coding partitions are skipped), coding partitions generator module 107 may be bypassed via switches 107a and 107b. In such examples, only a single level of partitioning may be performed. Such partitioning, where only a single level of partitioning exists, it may be described as prediction partitioning (as discussed) or coding partitioning or both. In various examples, such partitioning may be performed via prediction partitions generator module 105 (as discussed) or, as is discussed further herein, such partitioning may be performed via a k-d tree intra-prediction/coding partitioner module or a bi-tree intra-prediction/coding partitioner module implemented via coding partitions generator module 107.
In some examples, the partition prediction error data, if any, may not be significant enough to warrant encoding. In other examples, where it may be desirable to encode the partition prediction error data and the partition prediction error data is associated with inter-prediction or the like, coding partitions generator module 107 may determine coding partitions of the prediction partitions. In some examples, coding partitions generator module 107 may not be needed as the partition may be encoded without coding partitioning (e.g., as shown via the bypass path available via switches 107a and 107b). With or without coding partitioning, the partition prediction error data (which may subsequently be described as coding partitions in either event) may be transmitted to adaptive transform module 108 in the event the residuals or residual data require encoding. In some examples, prediction partitions generator module 105 and coding partitions generator module 107 may together be considered a partitioner subsystem of encoder 100. In various examples, coding partitions generator module 107 may operate on partition prediction error data, original pixel data, residual data, or wavelet data. Coding partitions generator module 107 may generate potential coding partitionings (for example, coding partitions) of, for example, partition prediction error data using bi-tree and/or k-d tree partitioning techniques or the like.
After the partitioning (after prediction partitions are formed for I-pictures, and coding partitions are formed for P- and F-pictures, and in some examples, the potential coding partitions), the partitions may be transformed using adaptive or fixed transforms with various block sizes via adaptive transform module 108 (also, in one form, referred to as the Adaptive Multi-size Rect Hybrid Parametric Haar Transform (HPHT)/Discrete Cosine Transform (DCT) unit). By one approach, the adaptive transform module 108 may perform forward HPHT or forward DCT on rectangular blocks. By one example, partition/block size as well as selected transforms (for example, adaptive or fixed, and HPHT or DCT) may be determined based on a rate distortion optimization (RDO) or other basis. In some examples, both the selected coding partitioning and/or the selected transform(s) may be determined based on a predetermined selection method based on coding partitions size or the like. For example, adaptive transform module 108 may include a first portion or component for performing a parametric transform to allow locally optimal transform coding of small to medium size blocks, and a second portion or component for performing globally stable, low overhead transform coding using a fixed transform, such as DCT or a picture based transform from a variety of transforms, including parametric transforms, or any other configuration. In some examples, for locally optimal transform coding, HPHT may be performed. In some examples, transforms may be performed on 2D blocks of rectangular sizes between about 4×4 pixels and 64×64 pixels, with actual sizes depending on a number of factors such as whether the transformed data is luma or chroma, or inter or intra, or if the determined transform used is PHT or DCT or the like.
For HPHT transform, small to medium block sizes are supported while for DCT transform a large number of block sizes are supported. For HPHT transform, some overhead is needed to identify the direction, either horizontal or vertical in which DCT is applied while the PHT is applied in the orthogonal direction, as well as the mode (at least for intra-coding where a mode can be based on decoded pixels or prediction difference pixels). The actual PHT transform basis used for transforming a particular block may be content adaptive as it depends on decoded neighboring pixels. Since both encoder and decoder require calculation of the same basis matrix, the complexity of the calculation is kept low by allowing a limited number of good transforms known (to both encoder and decoder) that one can select from.
As shown, the resultant transform coefficients may be transmitted to adaptive quantize module 109, while a quantizer adapter control 133 at the encode controller 103 performs analysis of content to come up with locally adaptive quantization parameters that are then represented by a multi-level map that can be efficiently coded and included in the bitstream. The computed quantizer set (qs, and a matrix applied to a coefficient block) may be used by the adaptive quantizer module 109 to perform scaling of the resultant transform coefficients. Further, any data associated with a parametric transform, as needed, may be transmitted to either adaptive quantize module 109 (if quantization is desired) or adaptive entropy encoder module 110. Also as shown in
As also shown in
The local decode loop 135 may begin at adaptive inverse quantize module 112. Adaptive inverse quantize module 112 may be configured to perform the opposite operation(s) of adaptive quantize module 109 such that an inverse scan may be performed and quantized coefficients may be de-scaled to determine transform coefficients. Such an adaptive quantize operation may be lossy, for example. As shown, the transform coefficients may be transmitted to an adaptive inverse transform module 113. Adaptive inverse transform module 113 may perform the inverse transform as that performed by adaptive transform module 108, for example, to generate residuals or residual values or partition prediction error data (or original data or wavelet data, as discussed) associated with coding partitions. In some examples, adaptive inverse quantize module 112 and adaptive inverse transform module 113 may together be considered a transform decoder subsystem of encoder 100.
As shown, the partition prediction error data (or the like) for P and F-pictures may be transmitted to optional coding partitions assembler 114. Coding partitions assembler 114 may assemble coding partitions into decoded prediction partitions as needed (as shown, in some examples, coding partitions assembler 114 may be skipped such as for I-picture tile partitioning, and via switches 114a and 114b such that decoded prediction partitions may have been generated at adaptive inverse transform module 113) to generate prediction partitions of prediction error data or decoded residual prediction partitions or the like. As shown, the decoded residual prediction partitions (inter or intra) may be added to predicted partitions (for example, prediction pixel data) at adder 115 to generate reconstructed prediction partitions. The reconstructed prediction partitions may be transmitted to prediction partitions assembler 116. Prediction partitions assembler 116 may assemble the reconstructed prediction partitions to generate reconstructed tiles or super-fragments. In some examples, coding partitions assembler module 114 and prediction partitions assembler module 116 may together be considered an un-partitioner sub-system of encoder 100.
The next set of steps involve filtering, and intermingling of filtering and prediction generation. Overall four types of filtering are shown. Specifically, in
More specifically, when the reconstructed tiles or super-fragments may be transmitted to blockiness analyzer and deblock filtering module 117, the blockiness analyzer and deblock filtering module 117 may deblock and dither the reconstructed tiles or super-fragments (or prediction partitions of tiles or super-fragments). The generated deblock and dither filter parameters may be used for the current filter operation and/or coded in bitstream 111 for use by a decoder, for example. The output of blockiness analyzer and deblock filtering module 117 may be transmitted to the quality analyzer and quality restoration filtering module 118. Quality analyzer and quality restoration filtering module 118 may determine QR filtering parameters (for example, for a QR decomposition) and use the determined parameters for filtering. The QR filtering parameters may also be coded in bitstream 111 for use by a decoder. In some examples, blockiness analyzer and deblock filtering module 117 and quality analyzer and quality restoration filtering module 118 may together be considered a filtering subsystem of encoder 100. In some examples, the output of quality analyzer and quality restoration filtering module 118 may be a final reconstructed frame that may be used for prediction for coding other frames (for example, the final reconstructed frame may be a reference frame or the like). Thus, as shown, the output of quality analyzer and quality restoration filtering module 118 may be transmitted to a multi-reference frame storage and frame selector (or multi reference control) 119 which also may be referred to as, or may include, the decoded picture storage or buffer. A dependency logic module 128 (also referred to, in one example, as dependency logic for mod multi ref pred in hierarchical picture group struct) may provide indices for listing the reference frames and the relationship among the frames such as frame dependencies, or more specifically partition dependencies, for proper ordering and use for the frames by the multi reference control 119 and when certain frames are to be selected for prediction of another frame. This may include providing the dependency logic for picture group structures such as multi-reference prediction, chain prediction, hierarchal structures, and/or other prediction techniques as described below.
Next, encoder 100 may perform inter- and/or intra-prediction operations. As shown in
The out-of-loop morphing analyzer 130 and the synthesizing analyzer 132 receive picture group structure data from the adaptive picture organizer 104 and communicate with the encoder controller 103 to form the morphing and synthesis parameters (mop, syp) and modified reference frames based on the non-quantized, non-decoded, original frame data. The formation of the modified reference frames and modification parameters from the out-of-loop morphing and synthesis analyzers 130 and 132 may be much faster than that provided through the decoder loop 135, and this is especially advantageous for real time encoding. However, the use of the modified frames and parameters to perform compensation at another location, such as by a decoder, should be performed by the in-loop morphing and synthesis generators 120 and 121 on the decoding loop side of the encoder so that the correct compensation can be repeated when reconstructing frames at the decoder. Thus, the resulting modification parameters from the out-of-loop analyzers 130 and 132 are used by the in-loop morphing and synthesizing generator 120 and 121 to form the modified reference frames and for motion estimation by the motion estimator 122 to compute motion vectors. Thus, the computed morphing and synthesis parameters (mop and syp) may be quantized/de-quantized and used (for example, by morphing generation module 120) to generate morphed reference frames that may be used by motion estimator module 122 for computing motion vectors for efficient motion (and characteristics) compensated prediction of a current frame. The synthesizing generation module 121 uses several types of synthesized frames including super resolution (SR) pictures, projected interpolation (PI) pictures, among others in which motion compensated prediction can result in even higher gains by determining motion vectors for efficient motion compensated prediction in these frames. The details for some examples to perform morphing or synthesis are provided below
Motion estimator module 122 may generate motion vector data based at least in part on morphed reference frame(s) and/or super resolution (SR) pictures and projected interpolation (PI) pictures along with the current frame. In some examples, motion estimator module 122 may be considered an inter-prediction module. For example, the motion vector data may be used for inter-prediction. If inter-prediction is applied, characteristics and motion filtering predictor module 123 may apply motion compensation as part of the local decode loop as discussed. Also, characteristics and motion filtering predictor module 123 may include adaptive precision (AP) filtering where filtering and prediction are intertwined. The filtering parameters (api) are coded and may be sent to the decoder via the bitstream 111.
Intra-prediction may be performed by intra-directional prediction analyzer and prediction generation module 124. Intra-directional prediction analyzer and prediction generation module 124 may be configured to perform spatial directional prediction and may use decoded neighboring partitions. In some examples, both the determination of direction and generation of prediction may be performed by intra-directional prediction analyzer and prediction generation module 124. In some examples, intra-directional prediction analyzer and prediction generation module 124 may be considered an intra-prediction module.
As shown in
In operation, some components of encoder 100 may operate as an encoder prediction subsystem. For example, such an encoder prediction subsystem of encoder 100 may include multi-reference frame storage and frame selector 119, in-loop morphing analyzer and generation module 120, in-loop synthesizing analyzer and generation module 121, motion estimator module 122, and/or characteristics and motion compensated precision adaptive filtering predictor module 123 as well as out-of-loop morphing analyzer 130 and synthesizing analyzer 132.
As will be discussed in greater detail below, in some implementations, such an encoder prediction subsystem of encoder 100 may incorporate a number of components and the combined predictions generated by these components in an efficient video coding algorithm. For example, proposed implementation of the NGV coder may include one or more of the following features: 1. Gain Compensation (e.g., explicit compensation for changes in gain/brightness in a scene); 2. Blur Compensation: e.g., explicit compensation for changes in blur/sharpness in a scene; 3. Dominant/Global Motion Compensation (e.g., explicit compensation for dominant motion in a scene); 4. Registration Compensation (e.g., explicit compensation for registration mismatches in a scene); 5. Super Resolution (e.g., explicit model for changes in resolution precision in a scene); 6. Projection (e.g., explicit model for changes in motion trajectory in a scene); the like, and/or combinations thereof.
For example, in such an encoder prediction subsystem of encoder 100, the output of quality analyzer and quality restoration filtering may be transmitted to multi-reference frame storage and frame selector 119. In some examples, the output of quality analyzer and quality restoration filtering may be a final reconstructed frame that may be used for prediction for coding other frames (e.g., the final reconstructed frame may be a reference frame or the like). In encoder 100, prediction operations may include inter- and/or intra-prediction. As shown, inter-prediction may be performed by one or more modules including morphing generation module 120, synthesizing generation module 121, and/or characteristics and motion compensated precision adaptive filtering predictor module 123.
As will be described in greater detail below, morphing generation module 120 may analyze a current frame to determine parameters for changes in gain, changes in dominant motion, changes in registration, and changes in blur with respect to a reference frame or frames with which it is to be coded. The determined morphing parameters may be quantized/de-quantized and used (e.g., by morphing generation module 120) to generate morphed reference frames. Such generated morphed reference frames may be stored in a buffer and may be used by motion estimator module 122 for computing motion vectors for efficient motion (and characteristics) compensated prediction of a current frame.
Similarly, synthesizing analyzer and generation module 121 may generate super resolution (SR) pictures and projected interpolation (PI) pictures or the like for determining motion vectors for efficient motion compensated prediction in these frames. Such generated synthesized reference frames may be stored in a buffer and may be used by motion estimator module 122 for computing motion vectors for efficient motion (and characteristics) compensated prediction of a current frame.
Accordingly, in such an encoder prediction subsystem of encoder 100, motion estimator module 122 may generate motion vector data based on morphed reference frame(s) and/or super resolution (SR) pictures and projected interpolation (PI) pictures along with the current frame. In some examples, motion estimator module 122 may be considered an inter-prediction module. For example, the motion vector data may be used for inter-prediction. If inter-prediction is applied, characteristics and motion filtering predictor module 123 may apply motion compensation as part of the local decode loop as discussed.
In operation, the proposed implementation of the NGV coder (e.g., encoder 100 and/or decoder 200) may use one or more of the above components besides the usual local motion compensation with respect to decoded past and/or future, picture/slices. As such the implementation does not mandate a specific solution for instance for dominant motion compensation, or for any other characteristics compensated reference frame generation.
The various signals and data items that may need to be sent to the decoder, i.e., pgst, ptyp, prp, pptn, cptn, modes, reftype, ethp, xmtyp, xmdir, xmmod, idir, mv, qs, mop, syp, ddi, qri, api, fii, quant coefficients and others may then be entropy encoded by adaptive entropy encoder 110 that may include different entropy coders collectively referred to as an entropy encoder subsystem. The adaptive entropy encoder 110 may be used to encode various types of control data/signals, parameters, modes and ref types, motion vectors, and transform coefficients. It is based on a generic class of low complexity entropy coders called adaptive variable length coders (vlc). The data to be entropy coded may be divided into several categories when convenient (seven in our case), and starting from generic vlc coders, specialized coders are developed for each category. While these control signals are illustrated as being associated with specific example functional modules of encoder 100 in
The final recon frames are saved in multi-reference frame storage and frame selector (also may be called decoded picture buffer) 210, and are used (or morphed) to create morphed pictures/local buffers (at morphed picture generator and buffer 211) depending on the applied, decoded mop parameters. Likewise synthesized picture and local buffers (at synthesized picture generation and buffer 212) are created by applying decoded syp parameters to multi-reference frame storage and frame selector 210 (or in other words, the reconstructed frames in the storage or buffer 210). A dependency logic 220 may hold the index for, and perform the indexing for, the stored frames in the multi-reference frame storage 210. The indexing may be used for prediction techniques such as multi-reference frames, chain prediction and/or hierarchal (or pyramid) frame structures, and/or others as described below. The morphed local buffers and synthesized frames are used for motion compensated prediction that uses adaptive precision (AP) filtering based on api parameters, and keeps either ¼ or ⅛ pel prediction depending on a decoded the ethp signal. In fact, a characteristics and motion compensated filtering predictor 213, depending on the mod, generates “inter” multi” “skip” or “auto” partitions while an intra-directional prediction generation module 214 generates “intra” partitions, and prediction modes selector 215, based on an encoder selected option, allows partition of the correct mode to pass through. Next, selective use of prediction fusion filter generation module (or Pred FI Filter Gen) 216 to filter and output the prediction is performed as needed as the second input to the adder.
The recon frames at the output of the quality filter generation module 209 (or Recon QR Filt Gen) are reordered (as F-pictures are out of order) by adaptive picture reorganizer (or Hierarchical Picture Group Structure Reorganizer) 217 in response to control parameters of ptyp and pgst, and further the output of this reorganizer undergoes optional processing in content post restorer 218 that is controlled by prp parameters sent by the encoder. This processing among other things may include deblocking and film grain addition.
More specifically, and as shown, decoder 200 may receive an input bitstream 201. In some examples, input bitstream 201 may be encoded via encoder 100 and/or via the encoding techniques discussed herein. As shown, input bitstream 201 may be received by an adaptive entropy decoder module 202. Adaptive entropy decoder module 202 may decode the various types of encoded data (e.g., overhead, motion vectors, transform coefficients, etc.). In some examples, adaptive entropy decoder 202 may use a variable length decoding technique. In some examples, adaptive entropy decoder 202 may perform the inverse operation(s) of adaptive entropy encoder module 110 discussed above.
The decoded data may be transmitted to adaptive inverse quantize module 203. Adaptive inverse quantize module 203 may be configured to inverse scan and de-scale quantized coefficients to determine transform coefficients. Such an adaptive quantize operation may be lossy, for example. In some examples, adaptive inverse quantize module 203 may be configured to perform the opposite operation of adaptive quantize module 109 (e.g., substantially the same operations as adaptive inverse quantize module 112). As shown, the transform coefficients (and, in some examples, transform data for use in a parametric transform) may be transmitted to an adaptive inverse transform module 204. Adaptive inverse transform module 204 may perform an inverse transform on the transform coefficients to generate residuals or residual values or partition prediction error data (or original data or wavelet data) associated with coding partitions. In some examples, adaptive inverse transform module 204 may be configured to perform the opposite operation of adaptive transform module 108 (e.g., substantially the same operations as adaptive inverse transform module 113). In some examples, adaptive inverse transform module 204 may perform an inverse transform based on other previously decoded data, such as, for example, decoded neighboring partitions. In some examples, adaptive inverse quantize module 203 and adaptive inverse transform module 204 may together be considered a transform decoder subsystem of decoder 200.
As shown, the residuals or residual values or partition prediction error data may be transmitted to coding partitions assembler 205. Coding partitions assembler 205 may assemble coding partitions into decoded prediction partitions as needed (as shown, in some examples, coding partitions assembler 205 may be skipped via switches 205a and 205b such that decoded prediction partitions may have been generated at adaptive inverse transform module 204). The decoded prediction partitions of prediction error data (e.g., prediction partition residuals) may be added to predicted partitions (e.g., prediction pixel data) at adder 206 to generate reconstructed prediction partitions. The reconstructed prediction partitions may be transmitted to prediction partitions assembler 207. Prediction partitions assembler 207 may assemble the reconstructed prediction partitions to generate reconstructed tiles or super-fragments. In some examples, coding partitions assembler module 205 and prediction partitions assembler module 207 may together be considered an un-partitioner subsystem of decoder 200.
The reconstructed tiles or super-fragments may be transmitted to deblock filtering module 208. Deblock filtering module 208 may deblock and dither the reconstructed tiles or super-fragments (or prediction partitions of tiles or super-fragments). The generated deblock and dither filter parameters may be determined from input bitstream 201, for example. The output of deblock filtering module 208 may be transmitted to a quality restoration filtering module 209. Quality restoration filtering module 209 may apply quality filtering based on QR parameters, which may be determined from input bitstream 201, for example. As shown in
As discussed, compensation due to prediction operations may include inter- and/or intra-prediction compensation. As shown, inter-prediction compensation may be performed by one or more modules including morphing generation module 211, synthesizing generation module 212, and characteristics and motion compensated filtering predictor module 213. Morphing generation module 211 may use de-quantized morphing parameters (e.g., determined from input bitstream 201) to generate morphed reference frames. Synthesizing generation module 212 may generate super resolution (SR) pictures and projected interpolation (PI) pictures or the like based on parameters determined from input bitstream 201. If inter-prediction is applied, characteristics and motion compensated filtering predictor module 213 may apply motion compensation based on the received frames and motion vector data or the like in input bitstream 201.
Intra-prediction compensation may be performed by intra-directional prediction generation module 214. Intra-directional prediction generation module 214 may be configured to perform spatial directional prediction and may use decoded neighboring partitions according to intra-prediction data in input bitstream 201.
As shown in
As discussed, the output of quality restoration filtering module 209 may be a final reconstructed frame. Final reconstructed frames may be transmitted to an adaptive picture re-organizer 217, which may re-order or re-organize frames as needed based on ordering parameters in input bitstream 201. Re-ordered frames may be transmitted to content post-restorer module 218. Content post-restorer module 218 may be an optional module configured to perform further improvement of perceptual quality of the decoded video. The improvement processing may be performed in response to quality improvement parameters in input bitstream 201 or it may be performed as standalone operation. In some examples, content post-restorer module 218 may apply parameters to improve quality such as, for example, an estimation of film grain noise or residual blockiness reduction (e.g., even after the deblocking operations discussed with respect to deblock filtering module 208). As shown, decoder 200 may provide display video 219, which may be configured for display via a display device (not shown).
In operation, some components of decoder 200 may operate as a decoder prediction subsystem. For example, such a decoder prediction subsystem of decoder 200 may include multi-reference frame storage and frame selector 210, dependency logic 220 to index the frames at the multi-reference frame storage and frame selector 210, morphing analyzer and generation module 211, synthesizing analyzer and generation module 212, and/or characteristics and motion compensated precision adaptive filtering predictor module 213.
As will be discussed in greater detail below, in some implementations, such a decoder prediction subsystem of decoder 200 may incorporate a number of components and the combined predictions generated by these components in an efficient video coding algorithm. For example, proposed implementation of the NGV coder may include one or more of the following features: 1. Gain Compensation (e.g., explicit compensation for changes in gain/brightness in a scene); 2. Blur Compensation: e.g., explicit compensation for changes in blur/sharpness in a scene; 3. Dominant/Global Motion Compensation (e.g., explicit compensation for dominant motion in a scene); 4. Registration Compensation (e.g., explicit compensation for registration mismatches in a scene); 5. Super Resolution (e.g., explicit model for changes in resolution precision in a scene); 6. Projection (e.g., explicit model for changes in motion trajectory in a scene); the like, and/or combinations thereof.
For example, in such a decoder prediction subsystem of decoder 200, the output of quality restoration filtering module may be transmitted to multi-reference frame storage and frame selector 210. In some examples, the output of quality restoration filtering module may be a final reconstructed frame that may be used for prediction for coding other frames (e.g., the final reconstructed frame may be a reference frame or the like). As discussed, compensation due to prediction operations may include inter- and/or intra-prediction compensation. As shown, inter-prediction compensation may be performed by one or more modules including morphing analyzer and generation module 211, synthesizing analyzer and generation module 212, and/or characteristics and motion compensated precision adaptive filtering predictor module 213.
As will be described in greater detail below, morphing analyzer and generation module 211 may use de-quantized morphing parameters (e.g., determined from input bitstream) to generate morphed reference frames. Such generated morphed reference frames may be stored in a buffer and may be used by characteristics and motion compensated precision adaptive filtering predictor module 213.
Similarly, synthesizing analyzer and generation module 212 may be configured to generate one or more types of synthesized prediction reference pictures such as super resolution (SR) pictures and projected interpolation (PI) pictures or the like based on parameters determined from input bitstream 201. Such generated synthesized reference frames may be stored in a buffer and may be used by motion compensated filtering predictor module 213.
Accordingly, in such a decoder prediction subsystem of decoder 200, in cases where inter-prediction is applied, characteristics and motion compensated filtering predictor module 213 may apply motion compensation based on morphed reference frame(s) and/or super resolution (SR) pictures and projected interpolation (PI) pictures along with the current frame.
In operation, the proposed implementation of the NGV coder (e.g., encoder 100 and/or decoder 200) may use one or more of the above components besides the usual local motion compensation with respect to decoded past and/or future, picture/slices. As such the implementation does not mandate a specific solution for instance for dominant motion compensation, or for any other characteristics compensated reference frame generation.
While
Further, it may be recognized that encoder 100 may be associated with and/or provided by a content provider system including, for example, a video content server system, and that output bitstream 111 may be transmitted or conveyed to decoders such as, for example, decoder 200 by various communications components and/or systems such as transceivers, antennae, network systems, and the like not depicted in
Partitioner subsystem 320a of encoder 300a may include two partitioning subsystems: prediction partitions generator module 105 that may perform analysis and partitioning for prediction, and coding partitions generator module 107 that may perform analysis and partitioning for coding. Another partitioning method may include adaptive picture organizer 104 which may segment pictures into regions or slices may also be optionally considered as being part of this partitioner.
Prediction encoder subsystem 330a of encoder 300a may include motion estimator 122 and characteristics and motion compensated filtering predictor 123 that may perform analysis and prediction of “inter” signal, and intra-directional prediction analyzer and prediction generation module 124 that may perform analysis and prediction of “intra” signal. Motion estimator 122 and characteristics and motion compensated filtering predictor 123 may allow for increasing predictability by first compensating for other sources of differences (such as gain, global motion, registration), followed by actual motion compensation. They may also allow for use of data modeling to create synthesized frames (super resolution, and projection) that may allow better predictions, followed by use of actual motion compensation in such frames.
Transform encoder subsystem 340a of encoder 300a may perform analysis to select the type and size of transform and may include two major types of components. The first type of component may allow for using parametric transform to allow locally optimal transform coding of small to medium size blocks; such coding however may require some overhead. The second type of component may allow globally stable, low overhead coding using a generic/fixed transform such as the DCT, or a picture based transform from a choice of small number of transforms including parametric transforms. For locally adaptive transform coding, PHT (Parametric Haar Transform) may be used. Transforms may be performed on 2D blocks of rectangular sizes between 4×4 and 64×64, with actual sizes that may depend on a number of factors such as if the transformed data is luma or chroma, inter or intra, and if the transform used is PHT or DCT. The resulting transform coefficients may be quantized, scanned and entropy coded.
Entropy encoder subsystem 360a of encoder 300a may include a number of efficient but low complexity components each with the goal of efficiently coding a specific type of data (various types of overhead, motion vectors, or transform coefficients). Components of this subsystem may belong to a generic class of low complexity variable length coding techniques, however, for efficient coding, each component may be custom optimized for highest efficiency. For instance, a custom solution may be designed for coding of “Coded/Not Coded” data, another for “Modes and Ref Types” data, yet another for “Motion Vector” data, and yet another one for “Prediction and Coding Partitions” data. Finally, because a very large portion of data to be entropy coded is “transform coefficient” data, multiple approaches for efficient handling of specific block sizes, as well as an algorithm that may adapt between multiple tables may be used.
Filtering encoder subsystem 350a of encoder 300a may perform analysis of parameters as well as multiple filtering of the reconstructed pictures based on these parameters, and may include several subsystems. For example, a first subsystem, blockiness analyzer and deblock filtering module 117 may deblock and dither to reduce or mask any potential block coding artifacts. A second example subsystem, quality analyzer and quality restoration filtering module 118, may perform general quality restoration to reduce the artifacts due to quantization operation in any video coding. A third example subsystem, which may include motion estimator 122 and characteristics and motion compensated filtering predictor module 123, may improve results from motion compensation by using a filter that adapts to the motion characteristics (motion speed/degree of blurriness) of the content. A fourth example subsystem, prediction fusion analyzer and filter generation module 126, may allow adaptive filtering of the prediction signal (which may reduce spurious artifacts in prediction, often from intra prediction) thereby reducing the prediction error which needs to be coded.
Encode controller module 103 of encoder 300a may be responsible for overall video quality under the constraints of given resources and desired encoding speed. For instance, in full RDO (Rate Distortion Optimization) based coding without using any shortcuts, the encoding speed for software encoding may be simply a consequence of computing resources (speed of processor, number of processors, hyperthreading, DDR3 memory etc.) availability. In such case, encode controller module 103 may be input every single combination of prediction partitions and coding partitions and by actual encoding, and the bitrate may be calculated along with reconstructed error for each case and, based on lagrangian optimization equations, the best set of prediction and coding partitions may be sent for each tile of each frame being coded. The full RDO based mode may result in best compression efficiency and may also be the slowest encoding mode. By using content analysis parameters from content pre-analyzer module 102 and using them to make RDO simplification (not test all possible cases) or only pass a certain percentage of the blocks through full RDO, quality versus speed tradeoffs may be made allowing speedier encoding. Up to now we have described a variable bitrate (VBR) based encoder operation. Encode controller module 103 may also include a rate controller that can be invoked in case of constant bitrate (CBR) controlled coding.
Lastly, pre-analyzer subsystem 310a of encoder 300a may perform analysis of content to compute various types of parameters useful for improving video coding efficiency and speed performance. For instance, it may compute horizontal and vertical gradient information (Rs, Cs), variance, spatial complexity per picture, temporal complexity per picture, scene change detection, motion range estimation, gain detection, prediction distance estimation, number of objects estimation, region boundary detection, spatial complexity map computation, focus estimation, film grain estimation etc. The parameters generated by preanalyzer subsystem 310a may either be consumed by the encoder or be quantized and communicated to decoder 200.
While subsystems 310a through 380a are illustrated as being associated with specific example functional modules of encoder 300a in
Entropy decoding subsystem 360b of decoder 300b may perform the inverse operation of the entropy encoder subsystem 360a of encoder 300a, i.e., it may decode various data (types of overhead, motion vectors, transform coefficients) encoded by entropy encoder subsystem 360a using a class of techniques loosely referred to as variable length decoding. Specifically, various types of data to be decoded may include “Coded/Not Coded” data, “Modes and Ref Types” data, “Motion Vector” data, “Prediction and Coding Partitions” data, and “Transform Coefficient” data.
Transform decoder subsystem 370b of decoder 300b may perform inverse operation to that of transform encoder subsystem 340a of encoder 300a. Transform decoder subsystem 370b may include two types of components. The first type of example component may support use of the parametric inverse PHT transform of small to medium block sizes, while the other type of example component may support inverse DCT transform for all block sizes. The PHT transform used for a block may depend on analysis of decoded data of the neighboring blocks. Output bitstream 111 and/or input bitstream 201 may carry information about partition/block sizes for PHT transform as well as in which direction of the 2D block to be inverse transformed the PHT may be used (the other direction uses DCT). For blocks coded purely by DCT, the partition/block sizes information may be also retrieved from output bitstream 111 and/or input bitstream 201 and used to apply inverse DCT of appropriate size.
Unpartitioner subsystem 380b of decoder 300b may perform inverse operation to that of partitioner subsystem 320a of encoder 300a and may include two unpartitioning subsystems, coding partitions assembler module 205 that may perform unpartitioning of coded data and prediction partitions assembler module 207 that may perform unpartitioning for prediction. Further if optional adaptive picture organizer module 104 is used at encoder 300a for region segmentation or slices, adaptive picture re-organizer module 217 may be needed at the decoder.
Prediction decoder subsystem 330b of decoder 300b may include characteristics and motion compensated filtering predictor module 213 that may perform prediction of “inter” signal and intra-directional prediction generation module 214 that may perform prediction of “intra” signal. Characteristics and motion compensated filtering predictor module 213 may allow for increasing predictability by first compensating for other sources of differences (such as gain, dominant motion, registration) or creation of synthesized frames (super resolution, and projection), followed by actual motion compensation.
Filtering decoder subsystem 350b of decoder 300b may perform multiple filtering of the reconstructed pictures based on parameters sent by encoder 300a and may include several subsystems. The first example subsystem, deblock filtering module 208, may deblock and dither to reduce or mask any potential block coding artifacts. The second example subsystem, quality restoration filtering module 209, may perform general quality restoration to reduce the artifacts due to quantization operation in any video coding. The third example subsystem, characteristics and motion compensated filtering predictor module 213, may improve results from motion compensation by using a filter that may adapt to the motion characteristics (motion speed/degree of blurriness) of the content. The fourth example subsystem, prediction fusion filtering module 216, may allow adaptive filtering of the prediction signal (which may reduce spurious artifacts in prediction, often from intra prediction) thereby reducing the prediction error which may need to be coded.
Post-restorer subsystem 390b of decoder 300b is an optional block that may perform further improvement of perceptual quality of decoded video. This processing can be done either in response to quality improvement parameters sent by encoder 100, or it can be standalone decision made at the post-restorer subsystem 390b. In terms of specific parameters computed at encoder 100 that can be used to improve quality at post-restorer subsystem 390b may be estimation of film grain noise and residual blockiness at encoder 100 (even after deblocking). As regards the film grain noise, if parameters can be computed and sent via output bitstream 111 and/or input bitstream 201 to decoder 200, then these parameters may be used to synthesize the film grain noise. Likewise, for any residual blocking artifacts at encoder 100, if they can be measured and parameters sent via output bitstream 111 and/or bitstream 201, post-restorer subsystem 390b may decode these parameters and may use them to optionally perform additional deblocking prior to display. In addition, encoder 100 also may have access to scene change, spatial complexity, temporal complexity, motion range, and prediction distance information that may help in quality restoration in post-restorer subsystem 390b.
While subsystems 330b through 390b are illustrated as being associated with specific example functional modules of decoder 300b in
The proposed implementation of the NGV coder (e.g., encoder 100 and/or decoder 200) may implement P-picture coding using a combination of Morphed Prediction References 428 through 438 (MR0 through 3) and/or Synthesized Prediction References 412 and 440 through 446 (S0 through S3, MR4 through 7). NGV coding involves use of three picture types referred to as I-pictures, P-pictures, and FB-pictures. In the illustrated example, the current picture to be coded (a P-picture) is shown at time t=4. During coding, the proposed implementation of the NGV coder (e.g., encoder 100 and/or decoder 200) may use one or more of four previously decoded references R0412, R1414, R2416, and R3418. Unlike other solutions that may simply use these references directly for prediction, the proposed implementation of the NGV coder (e.g., encoder 100 and/or decoder 200) may generate modified (morphed or synthesized) references from such previously decoded references and then use motion compensated coding based at least in part on such generated modified (morphed or synthesized) references.
As will be described in greater detail below, in some examples, the proposed implementation of the NGV coder (e.g., encoder 100 and/or decoder 200) may incorporate a number of components and the combined predictions generated by these components in an efficient video coding algorithm. For example, proposed implementation of the NGV coder may include one or more of the following features: 1. Gain Compensation (e.g., explicit compensation for changes in gain/brightness in a scene); 2. Blur Compensation: e.g., explicit compensation for changes in blur/sharpness in a scene; 3. Dominant/Global Motion Compensation (e.g., explicit compensation for dominant motion in a scene); 4. Registration Compensation (e.g., explicit compensation for registration mismatches in a scene); 5. Super Resolution (e.g., explicit model for changes in resolution precision in a scene); 6. Projection (e.g., explicit model for changes in motion trajectory in a scene); the like, and/or combinations thereof.
In the illustrated example, if inter-prediction is applied, a characteristics and motion filtering predictor module may apply motion compensation to a current picture 410 (e.g., labeled in the figure as P-pic (curr)) as part of the local decode loop. In some instances, such motion compensation may be based at least in part on future frames (not shown) and/or previous frame R0412 (e.g., labeled in the figure as R0), previous frame R1414 (e.g., labeled in the figure as R1), previous frame R2416 (e.g., labeled in the figure as R2), and/or previous frame R3418 (e.g., labeled in the figure as R3).
For example, in some implementations, prediction operations may include inter- and/or intra-prediction. Inter-prediction may be performed by one or more modules including a morphing analyzer and generation module and/or a synthesizing analyzer and generation module. Such a morphing analyzer and generation module may analyze a current picture to determine parameters for changes in blur 420 (e.g., labeled in the figure as Blur par), changes in gain 422 (e.g., labeled in the figure as Gain par and explained in detail below), changes in registration 424 (e.g., labeled in the figure as Reg par), and changes in dominant motion 426 (e.g., labeled in the figure as Dom par), or the like with respect to a reference frame or frames with which it is to be coded.
The determined morphing parameters 420, 422, 424, and/or 426 may be used to generate morphed reference frames. Such generated morphed reference frames may be stored and may be used for computing motion vectors for efficient motion (and characteristics) compensated prediction of a current frame. In the illustrated example, determined morphing parameters 420, 422, 424, and/or 426 may be used to generate morphed reference frames, such as blur compensated morphed reference frame 428 (e.g., labeled in the figure as MR3b), gain compensated morphed reference frame 430 (e.g., labeled in the figure as MR2g), gain compensated morphed reference frame 432 (e.g., labeled in the figure as MR1g), registration compensated morphed reference frame 434 (e.g., labeled in the figure as MR1r), dominant motion compensated morphed reference frame 436 (e.g., labeled in the figure as MR0d), and/or registration compensated morphed reference frame 438 (e.g., labeled in the figure as MR0r), the like or combinations thereof, for example.
Similarly, a synthesizing analyzer and generation module may generate super resolution (SR) pictures 440 (e.g., labeled in the figure as S0 (which is equal to previous frame R0412), S1, S2, S3) and projected interpolation (PI) pictures 442 (e.g., labeled in the figure as PE) or the like for determining motion vectors for efficient motion compensated prediction in these frames. Such generated synthesized reference frames may be stored and may be used for computing motion vectors for efficient motion (and characteristics) compensated prediction of a current frame.
Additionally or alternatively, the determined morphing parameters 420, 422, 424, and/or 426 may be used to morph the generated synthesis reference frames super resolution (SR) pictures 440 and/or projected interpolation (PI) pictures 442. For example, a synthesizing analyzer and generation module may generate morphed registration compensated super resolution (SR) pictures 444 (e.g., labeled in the figure as MR4r, MR5r, and MR6r) and/or morphed registration compensated projected interpolation (PI) pictures 446 (e.g., labeled in the figure as MR7r) or the like from the determined registration morphing parameter 424. Such generated morphed and synthesized reference frames may be stored and may be used for computing motion vectors for efficient motion (and characteristics) compensated prediction of a current frame.
In some implementations, changes in a set of characteristics (such as gain, blur, dominant motion, registration, resolution precision, motion trajectory, the like, or combinations thereof, for example) may be explicitly computed. Such a set of characteristics may be computed in addition to local motion. In some cases previous and next pictures/slices may be utilized as appropriate; however, in other cases such a set of characteristics may do a better job of prediction from previous picture/slices. Further, since there can be error in any estimation procedure, (e.g., from multiple past or multiple past and future pictures/slices) a modified reference frame associated with the set of characteristics (such as gain, blur, dominant motion, registration, resolution precision, motion trajectory, the like, or combinations thereof, for example) may be selected that yields the best estimate. Thus, the proposed approach that utilizes modified reference frames associated with the set of characteristics (such as gain, blur, dominant motion, registration, resolution precision, motion trajectory, the like, or combinations thereof, for example) may explicitly compensate for differences in these characteristics. The proposed implementation may address the problem of how to improve the prediction signal, which in turn allows achieving high compression efficiency in video coding.
As discussed, in some examples, inter-prediction may be performed. In some examples, up to 4 decoded past and/or future pictures and several morphing/synthesis predictions may be used to generate a large number of reference types (e.g., reference pictures). For instance in ‘inter’ mode, up to nine reference types may be supported in P-pictures, and up to ten reference types may be supported for F/B-pictures. Further, ‘multi’ mode may provide a type of inter prediction mode in which instead of 1 reference picture, two reference pictures may be used and P- and F/B-pictures respectively may allow 3, and up to 8 reference types. For example, prediction may be based on a previously decoded frame generated using at least one of a morphing technique or a synthesizing technique. In such examples, the bitstream may include a frame reference, morphing parameters, or synthesizing parameters associated with the prediction partition.
Some of the morphing and synthesis techniques other than the dominant motion compensation (described in more detail below) are as follows.
Gain Compensation
One type of morphed prediction used by NGV coding is gain compensated prediction, and includes detecting and estimating the gain and/or offset luminance values, parameterizing them, using them for compensation of gain/offset at the encoder, transmitting them to the decoder, and using them at the decoder for gain compensation by replicating the gain compensation process at the encoder.
By one detailed example, often in video scenes, frame to frame differences are caused not only due to movement of objects but also due to changes in gain/brightness. Sometimes such changes in brightness can be global due to editing effects such as a fade-in, a fade-out, or due to a crossfade. However, in many more cases, such changes in brightness are local for instance due to flickering lights, camera flashes, explosions, colored strobe lights in a dramatic or musical performance, etc.
The compensation of interframe changes in brightness, whether global or local, can potentially improve compression efficiency in video coding. However, the brightness change parameters (gain and offset) are applied both at a video encoder and a decoder so that both should be efficiently communicating with low bit-cost from encoder to decoder via the bitstream and the processing complexity for the decoder should be minimized. In the past, only techniques for global brightness change have been disclosed, but local compensation in brightness changes have not been successfully addressed.
The following equation relates brightness of a pixel st(i,j) at (i,j) location in frame ‘t’ to brightness of a pixel at the same location (i,j) in a previous frame ‘t−1’, with ‘a’ and ‘b’ being the gain and offset factors. Motion is assumed to be small and only the brightness changes are modeled.
s
t(i,j)=a×st-1(i,j)+b (1)
Taking the expected value of st(i,j) and (st2(i,j)), and following a method of equating first and second moments of current frame and the previous frame, the value of gain ‘a’ and offset ‘b’ can then be calculated as:
Once ‘a’ and ‘b’ are calculated as per equation (2), they are quantized (for efficient transmission), encoded and sent to the decoder. At the decoder, decoded dequantized values of ‘a’, and ‘b’ are put back into equation (1), and using decoded values of pixels in the previous frame, a gain compensated modified version of a previous reference frame is calculated that is lower in error than the original previous frame, and is then used for generating (gain compensated) motion compensated prediction. To the (inverse transformed, and dequantized) decoded prediction error blocks, the corresponding predictions from modified previous reference frames are added to generate the final decoded frame (or blocks of the frame).
For local motion compensation, instead of a single set of (a, b) parameters, multiple sets of parameters are computed and transmitted along with the map of which portion of the frame corresponds to which parameters, and to the decoder and used for gain compensation as described.
Blur/Registration Compensation
By one detailed example, methods for compensation of Registration and Blur are described below although the terms can be used interchangeably.
Registration Compensation:
A stationary video camera imaging a scene might still result in shaky or unstable video that differs frame to frame due to environmental factors (such as wind), vibrations from nearby objects, a shaky hand, or a jittery capture process, rather than global movement of the scene or motion of large objects in the scene. This results in frame to frame registration differences, the compensation of which (in addition to other forms of compensation such as gain, global/dominant motion, and local motion compensation) may result in improvement of compression efficiency of video coding.
For computing registration parameters between a current frame and a previous reference frame, Wiener filtering can be employed. Let x(n) be the input signal, y(n) be the output, and h(n) represent filter coefficients.
In matrix notation, h is the vector of filter coefficients. The cross-correlation row vector (between source frame and reference frame):
R
dx
=E[d(n)x(n)T] (6)
R
xx
=E[x(n)x(n)T] (7)
The Wiener Hopf equation to solve for h as then as follows. The Wiener Hopf equation determines optimum filter coefficients in mean square error, and the resulting filter is called the ‘wiener’ filter.
h=R
xx
−1
R
dx (8)
A fast camera pan of a scene may, due to charge integration, result in blurry image. Further, even if a camera is still, or in motion, if a scene involves fast moving objects, for instance football players in a football game, the objects can appear blurry as the temporal resolution of the imaging is not sufficient. In both of the aforementioned cases, compensation of blur prior to or in conjunction with other forms of compensation, may improve compression efficiency of video coding.
For motion blur estimation, a Lucy-Richardson method can be used. It is an iterative algorithm for successively computing reduced blur frame (X) at iteration i, from Y the source frame, using B, the blur operator (blur frame using estimated blur vectors) and B* an adjoint operator. The operator B* can be roughly thought of as the same as B as B* can be replaced by B resulting in roughly the same visual quality.
Super Resolution Synthesis
Referring to
For sequences where the movement is slow and content is fairly detailed (many edges, texture, and so forth), the ability to generate super resolution frames for use in prediction can provide greater motion compensation accuracy, and thereby permit a higher degree of compression. As shown in
Projected Interpolation Synthesis
A picture sequence such as frame sequence 400, may also be used to illustrate the principle of generation and use of projected interpolation frames (PI-pictures) shown as frame PE 442 on
For instance, for F-pictures at a time ‘n+1’, a PI-picture is generated co-located at this time using anchor or reference frames at times ‘n’ and ‘n+2’. Likewise for F-pictures at times, ‘n+3’, and ‘n+4’, corresponding PI-pictures can be generated using anchor frames at times ‘n+2’ and ‘n+5’. This process may repeat for each future F-picture as a PI-picture is synthesized to correspond in time to each F-picture. The corresponding synthesized PI-pictures can then be used as a third reference in the same or similar way the two reference anchors were going to be used for prediction. Some prediction partitions may use prediction references directly while others may use them implicitly such as to generate bi-prediction. Thus, synthesized PI-pictures can be used for prediction, instead of the original F-pictures, with multi-reference prediction and with two reference anchors.
Turning now to the system to implement these modifications to reference frames, and as mentioned previously, with ever increasing resolution of video to be compressed and expectation of high video quality, the corresponding bitrate/bandwidth required for coding using existing video coding standards such as H.264 or even evolving standards such as H.265/HEVC, is relatively high. The aforementioned standards use expanded forms of traditional approaches to implicitly address the insufficient compression/quality problem, but often the results are limited.
The proposed implementation improves video compression efficiency by improving interframe prediction, which in turn reduces interframe prediction difference (error signal) that needs to be coded. The less the amount of interframe prediction difference to be coded, the less the amount of bits required for coding, which effectively improves the compression efficiency as it now takes less bits to store or transmit the coded prediction difference signal. Instead of being limited to motion predictions only, the proposed NCV codec may be highly adaptive to changing characteristics (such as gain, blur, dominant motion, registration, resolution precision, motion trajectory, the like, or combinations thereof, for example) of the content by employing, in addition or in the alternative to motion compensation, approaches to explicitly compensate for changes in the characteristics of the content. Thus, by explicitly addressing the root cause of the problem the NGV codec may address a key source of limitation of standards based codecs, thereby achieving higher compression efficiency.
This change in interframe prediction output may be achieved due to ability of the proposed NCV codec to compensate for a wide range of reasons for changes in the video content. Typical video scenes vary from frame to frame due to many local and global changes (referred to herein as characteristics). Besides local motion, there are many other characteristics that are not sufficiently addressed by current solutions that may be addressed by the proposed implementation.
The proposed implementation may explicitly compute changes in a set of characteristics (such as gain, blur, dominant motion, registration, resolution precision, motion trajectory, the like, or combinations thereof, for example) in addition to local motion, and thus may do a better job of prediction from previous picture/slices than only using local motion prediction from previous and next pictures/slices. Further, since there can be error in any estimation procedure, from multiple past or multiple past and future pictures/slices the NGV coder may choose the frame that yields the best by explicitly compensating for differences in various characteristics.
In operation, the proposed implementation of the NGV coder (e.g., encoder 100 and/or decoder 200) may operate so that prediction mode and/or reference type data may be defined using symbol-run coding or a codebook or the like. The prediction mode and/or reference type data may be transform encoded using content adaptive or discrete transform in various examples to generate transform coefficients. Also as discussed, data associated with partitions (e.g., the transform coefficients or quantized transform coefficients), overhead data (e.g., indicators as discussed herein for transform type, adaptive transform direction, and/or a transform mode), and/or data defining the partitions and so on may be encoded (e.g., via an entropy encoder) into a bitstream. The bitstream may be communicated to a decoder, which may use the encoded bitstream to decode video frames for display. On a local basis (such as block-by-block within a macroblock or a tile, or on a partition-by-partition within a tile or a prediction unit, or fragments within a superfragment or region) the best mode may be selected for instance based at least in part on Rate Distortion Optimization (RDO) or based at least in part on pre-analysis of video, and the identifier for the mode and needed references may be encoded within the bitstream for use by the decoder.
As explained above, various prediction modes are allowed in P- and F-pictures and are exemplified below, along with how they relate to the reference types. Both the P-picture and F-picture tiles are partitioned into smaller units, and a prediction mode from among “skip”, “auto”, “inter”, and “multi”, is assigned to each partition of a tile. The entire list of modes in Table 1 also includes ‘intra’ that refers to spatial prediction from neighboring blocks as compared to temporal motion compensated prediction. The “split” mode refers to a need for further division or further partitioning. For partitions that use “inter” or “multi” mode, further information about the used reference is needed and is shown for P-pictures in Table 2(a) and Table 2(b) respectively, while for F-pictures, in Table 3(a) and Table 3(b) respectively.
Prediction modes and reference types analyzer 125 (
Tables 1 through 3(d), shown below, illustrate one example of codebook entries for a current frame (curr_pic) being, or that will be, reconstructed. A full codebook of entries may provide a full or substantially full listing of all possible entries and coding thereof. In some examples, the codebook may take into account constraints as described above. In some examples, data associated with a codebook entry for prediction modes and/or reference types may be encoded in a bitstream for use at a decoder as discussed herein.
where table 2(b) is directed to a specific combination of references including a past reference without parameters and one of the references on the table as indicated by the table heading.
where proj F refers to PI, and line 8, by one example, includes two optional references.
where Dir refers to a sub-mode that is a fixed, or partially fixed, combination of references for multi-mode for F-frames, such that Dir 0 above, and Dir 1 and Dir 2 below, each refer to a combination of references. Thus, as shown in Table 3(b), Dir 0 may refer to a combination of a past reference (which may be a particular reference at a particular time (reference 3 at n+2 for example) and combined with one of the references from the table. Dir on the tables below are similar and as explained in the heading of the table.
Specific to dominant motion compensation, “inter” mode of P-pictures supports a reference type called MR0d (morphed reference 0 dominant motion), and for “inter” mode of F-pictures, supported reference types include MR0d and MR3d (morphed reference 3 dominant motion). These codes are explained further below. Further, in “multi” mode, MR3d is supported as one of the two references used for a single current frame. Besides “inter” and “multi”, DMC also may be used in “auto” mode of NGV. A summary of modes and reference type combinations where DMC is invoked are as follows.
Use the dominant motion compensated reference frame prediction:
Use blended prediction of multiple dominant motion compensated reference frames:
Dominant motion compensated reference with differential translational motion vector for prediction:
Dominant motion compensated reference with differential translational motion vector for prediction, blended with another reference frame:
As shown, the output of quality analyzer and quality restoration filtering may be transmitted to decoded picture buffer 119. In some examples, the output of quality analyzer and quality restoration filtering may be a final reconstructed frame that may be used for prediction for coding other frames (e.g., the final reconstructed frame may be a reference frame or the like). In encoder 600, prediction operations may include inter- and/or intra-prediction. As shown in
Morphing analyzer and generation module 120 may include a morphing types analyzer (MTA) and a morphed pictures generator (MPG) 610 as well as a morphed prediction reference (MPR) buffer 620. Morphing types analyzer (MTA) and a morphed pictures generator (MPG) 610 may analyze a current picture to determine parameters for changes in gain, changes in dominant motion, changes in registration, and changes in blur with respect to a reference frame or frames with which it is to be coded. The determined morphing parameters may be quantized/de-quantized and used (e.g., by morphing analyzer and generation module 120) to generate morphed reference frames. Such generated morphed reference frames may be stored in morphed prediction reference (MPR) buffer 620 and may be used by motion estimator module 122 for computing motion vectors for efficient motion (and characteristics) compensated prediction of a current frame.
Synthesizing analyzer and generation module 121 may include a synthesis types analyzer (STA) and synthesized pictures generator (SPG) 630 as well as a synthesized prediction reference (SPR) buffer 640. Synthesis types analyzer (STA) and synthesized pictures generator 630 may generate super resolution (SR) pictures and projected interpolation (PI) pictures or the like for determining motion vectors for efficient motion compensated prediction in these frames. Such generated synthesized reference frames may be stored in synthesized prediction reference (SPR) buffer 640 and may be used by motion estimator module 122 for computing motion vectors for efficient motion (and characteristics) compensated prediction of a current frame.
Motion estimator module 122 may generate motion vector data based at least in part on morphed reference frame(s) and/or super resolution (SR) pictures and projected interpolation (PI) pictures along with the current frame. In some examples, motion estimator module 122 may be considered an inter-prediction module. For example, the motion vector data may be used for inter-prediction. If inter-prediction is applied, characteristics and motion filtering predictor module 123 may apply motion compensation as part of the local decode loop as discussed.
As shown, the output of quality restoration filtering module may be transmitted to decoded picture buffer (or frame selector control) 210. In some examples, the output of quality restoration filtering module may be a final reconstructed frame that may be used for prediction for coding other frames (e.g., the final reconstructed frame may be a reference frame or the like). As discussed, compensation due to prediction operations may include inter- and/or intra-prediction compensation. As shown, inter-prediction compensation may be performed by one or more modules including morphing analyzer and generation module 211, synthesizing analyzer and generation module 212, and/or characteristics and motion compensated precision adaptive filtering predictor module 213.
Morphing analyzer and generation module 211 may include a morphed pictures generator (MPG) 710 as well as a morphed prediction reference (MPR) buffer 720. Morphed pictures generator (MPG) 710 may use de-quantized morphing parameters (e.g., determined from input bitstream) to generate morphed reference frames. Such generated morphed reference frames may be stored in morphed prediction reference (MPR) buffer 720 and may be used by characteristics and motion compensated precision adaptive filtering predictor module 213.
Synthesizing analyzer and generation module 212 may include a synthesized pictures generator (SPG) 730 as well as a synthesized prediction reference (SPR) buffer 740. Synthesized pictures generator 730 may be configured to generate one or more types of synthesized prediction reference pictures such as super resolution (SR) pictures and projected interpolation (PI) pictures or the like based at least in part on parameters determined from input bitstream 201. Such generated synthesized reference frames may be stored in synthesized prediction reference (SPR) buffer 740 and may be used by motion compensated filtering predictor module 213.
If inter-prediction is applied, characteristics and motion compensated filtering predictor module 213 may apply motion compensation based at least in part on morphed reference frame(s) and/or super resolution (SR) pictures and projected interpolation (PI) pictures along with the current frame.
Referring to
As shown, the output of quality analyzer and quality restoration filtering may be transmitted to decoded picture buffer 119. In some examples, the output of quality analyzer and quality restoration filtering may be a final reconstructed frame that may be used for prediction for coding other frames (e.g., the final reconstructed frame may be a reference frame or the like). In encoder 800, prediction operations may include inter- and/or intra-prediction. As shown in
Morphing analyzer and generation module 120 may include a morphing types analyzer (MTA) and a morphed pictures generator (MPG) 610 as well as a morphed prediction reference (MPR) buffer 620. Morphing types analyzer (MTA) and a morphed pictures generator (MPG) 610 may be configured to analyze and/or generate one or more types of modified prediction reference pictures.
For example, morphing types analyzer (MTA) and a morphed pictures generator (MPG) 610 may include Gain Estimator and Compensated Prediction Generator 805, Blur Estimator and Compensated Prediction Generator 810, Dominant Motion Estimator and Compensated Prediction Generator 815, Registration Estimator and Compensated Prediction Generator 820, the like and/or combinations thereof. Gain Estimator and Compensated Prediction Generator 805 may be configured to analyze and/or generate morphed prediction reference pictures that are adapted to address changes in gain. Blur Estimator and Compensated Prediction Generator 810 may be configured to analyze and/or generate morphed prediction reference pictures that are adapted to address changes in blur. Global Motion Estimator and Compensated Prediction Generator 815 may be configured to analyze and/or generate morphed prediction reference pictures that are adapted to address changes in dominant motion. Specifically, the global motion estimator and compensated prediction generator 815 is used for computation of global motion parameters (dp) and applying them on a picture from the DPR buffers 119 to generate a GMC Morphed Reference Picture that is stored in one of the MPR Picture Buffers (Local/Picture Buffers for Dominant Motion Compensated Prediction). The output of that is used for block motion estimation and compensation. Registration Estimator and Compensated Prediction Generator 820 may be configured to analyze and/or generate morphed prediction reference pictures that are adapted to address changes in registration.
Morphing types analyzer (MTA) and a morphed pictures generator (MPG) 610 may store such generated morphed reference frames in morphed prediction reference (MPR) buffer 620. For example, morphed prediction reference (MPR) buffer 620 may include Gain Compensated (GC) Picture/s Buffer 825, Blur Compensated (BC) Picture/s Buffer 830, Dominant Motion Compensated (DC) Picture/s Buffer 835, Registration Compensated (RC) Picture/s Buffer 840, the like and/or combinations thereof. Gain Compensated (GC) Picture/s Buffer 825 may be configured to store morphed reference frames that are adapted to address changes in gain. Blur Compensated (BC) Picture/s Buffer 830 may be configured to store morphed reference frames that are adapted to address changes in blur. Dominant Motion Compensated (DC) Picture/s Buffer 835 may be configured to store morphed reference frames that are adapted to address changes in dominant motion. Registration Compensated (RC) Picture/s Buffer 840 may be configured to store morphed reference frames that are adapted to address changes in registration.
Synthesizing analyzer and generation module 121 may include a synthesis types analyzer (STA) and synthesized pictures generator 630 as well as a synthesized prediction reference (SPR) buffer 640. Synthesis types analyzer (STA) and synthesized pictures generator 530 may be configured to analyze and/or generate one or more types of synthesized prediction reference pictures. For example, synthesis types analyzer (STA) and synthesized pictures generator 630 may include Super Resolution Filter Selector & Prediction Generator 845, Projection Trajectory Analyzer & Prediction Generator 850, the like and/or combinations thereof. Super Resolution Filter Selector & Prediction Generator 845 may be configured to analyze and/or generate a super resolution (SR) type of synthesized prediction reference pictures. Projection Trajectory Analyzer & Prediction Generator 850 may be configured to analyze and/or generate a projected interpolation (PI) type of synthesized prediction reference pictures.
Synthesis types analyzer (STA) and synthesized pictures generator 630 may generate super resolution (SR) pictures and projected interpolation (PI) pictures or the like for efficient motion compensated prediction in these frames. Such generated synthesized reference frames may be stored in synthesized prediction reference (SPR) buffer 640 and may be used by motion estimator module 122 for computing motion vectors for efficient motion (and characteristics) compensated prediction of a current frame.
For example, synthesized prediction reference (SPR) buffer 640 may include Super Resolution (SR) Picture Buffer 855, Projected Interpolation (PI) Picture Buffer 860, the like and/or combinations thereof. Super Resolution (SR) Picture Buffer 855 may be configured to store synthesized reference frames that are generated for super resolution (SR) pictures. Projected Interpolation (PI) Picture Buffer 860 may be configured to store synthesized reference frames that are generated for projected interpolation (PI) pictures.
Motion estimator module 122 may generate motion vector data based on morphed reference frame(s) and/or super resolution (SR) pictures and projected interpolation (PI) pictures along with the current frame. In some examples, motion estimator module 122 may be considered an inter-prediction module. For example, the motion vector data may be used for inter-prediction. If inter-prediction is applied, characteristics and motion filtering predictor module 123 may apply motion compensation as part of the local decode loop as discussed.
The prediction mode analyzer 125 (or Pred Modes & Ref Types Analyzer & Selector), and as explained above, chooses on a local (block, tile or partition) basis the best prediction from among various type of inter modes and intra mode. Here the term inter is being used in generality and includes ‘inter’ mode, ‘multi’ mode, ‘auto’ mode and ‘skip’ modes. The chosen mode (and sub-mode if applicable), morphing or synthesis parameters (dp, gp, rp, sp, pp), reference info, and motion (mv, Δmv) and other data is entropy coded as explained above and sent as part of an encoded bitstream to the decoder.
As shown, the output of quality restoration filtering module may be transmitted to decoded picture buffer 210. In some examples, the output of quality restoration filtering module may be a final reconstructed frame that may be used for prediction for coding other frames (e.g., the final reconstructed frame may be a reference frame or the like). As discussed, compensation due to prediction operations may include inter- and/or intra-prediction compensation. As shown, inter-prediction compensation may be performed by one or more modules including morphing analyzer and generation module 211, synthesizing analyzer and generation module 212, and/or characteristics and motion compensated precision adaptive filtering predictor module 213.
Morphing generation module 212 may include a morphed pictures generator (MPG) 710 as well as a morphed prediction reference (MPR) buffer 720. Morphed pictures generator (MPG) 710 may use de-quantized morphing parameters (e.g., determined from input bitstream) to generate morphed reference frames. For example, morphed pictures generator (MPG) 710 may include Gain Compensated Prediction Generator 905, Blur Compensated Prediction Generator 910, Dominant Motion Compensated Prediction Generator 915, Registration Compensated Prediction Generator 920, the like and/or combinations thereof. Gain Compensated Prediction Generator 905 may be configured to generate morphed prediction reference pictures that are adapted to address changes in gain as described in greater detail below. Blur Compensated Prediction Generator 910 may be configured to generate morphed prediction reference pictures that are adapted to address changes in blur. Dominant Motion Compensated Prediction Generator 915 may be configured to generate morphed prediction reference pictures that are adapted to address changes in dominant motion. Registration Compensated Prediction Generator 920 may be configured to generate morphed prediction reference pictures that are adapted to address changes in registration.
Morphed pictures generator (MPG) 710 may store such generated morphed reference frames in morphed prediction reference (MPR) buffer 720. For example, morphed prediction reference (MPR) buffer 720 may include Gain Compensated (GC) Picture/s Buffer 925, Blur Compensated (BC) Picture/s Buffer 930, Dominant Motion Compensated (DC) Picture/s Buffer 935, Registration Compensated (RC) Picture/s Buffer 940, the like and/or combinations thereof. Gain Compensated (GC) Picture/s Buffer 925 may be configured to store morphed reference frames that are adapted to address changes in gain. Blur Compensated (BC) Picture/s Buffer 930 may be configured to store morphed reference frames that are adapted to address changes in blur. Dominant Motion Compensated (DC) Picture/s Buffer 935 may be configured to store morphed reference frames that are adapted to address changes in dominant motion. Registration Compensated (RC) Picture/s Buffer 940 may be configured to store morphed reference frames that are adapted to address changes in registration.
Synthesizing generation module 212 may include a synthesized pictures generator 630 as well as a synthesized prediction reference (MPR) buffer 740. Synthesized pictures generator 730 may be configured to generate one or more types of synthesized prediction reference pictures such as super resolution (SR) pictures and projected interpolation (PI) pictures or the like based on parameters determined from input bitstream 201. Such generated synthesized reference frames may be stored in synthesized prediction reference (SPR) buffer 740 and may be used by motion compensated filtering predictor module 213. For example, synthesized pictures generator 730 may include Super Resolution Picture Generator 945, Projection Trajectory Picture Generator 950, the like and/or combinations thereof. Super Resolution Picture Generator 945 may be configured to generate a super resolution (SR) type of synthesized prediction reference pictures. Projection Trajectory Picture Generator 950 may be configured to generate a projected interpolation (PI) type of synthesized prediction reference pictures.
Synthesized pictures generator 730 may generate super resolution (SR) pictures and projected interpolation (PI) pictures or the like for efficient motion compensated prediction in these frames. Such generated synthesized reference frames may be stored in synthesized prediction reference (SPR) buffer 740 and may be used by characteristics and motion compensated filtering predictor module 213 for efficient motion (and characteristics) compensated prediction of a current frame.
For example, synthesized prediction reference (SPR) buffer 740 may include Super Resolution (SR) Picture Buffer 955, Projected Interpolation (PI) Picture Buffer 960, the like and/or combinations thereof. Super Resolution (SR) Picture Buffer 955 may be configured to store synthesized reference frames that are generated for super resolution (SR) pictures. Projected Interpolation (PI) Picture Buffer 960 may be configured to store synthesized reference frames that are generated for projected interpolation (PI) pictures.
If inter-prediction is applied, characteristics and motion compensated filtering predictor module 213 may apply motion compensation based on morphed reference frame(s) and/or super resolution (SR) pictures and projected interpolation (PI) pictures along with the current frame.
Dominant Motion Compensation
Referring to
One of the limitations of global motion compensation (GMC) included in the MPEG-4 standard is that the computed GMC parameters may not provide good prediction due to various reasons including large distance between prediction frames, uncovered background, mixing of global and local motion of objects, and simplistic interpolation. With no way for correction of computed GMC parameters, the only alternative existed was to enable or disable use of GMC on a local basis. More specifically, it was determined whether or not to use GMC on a block-by-block basis. This is an overhead-expensive process.
In comparison, NGV video coding introduces the principle of correction of GMC prediction by introducing a correction vector. By one example, an original or current video picture 1002 to be coded (on the right) has a foreground object 1004 (a large star shaped object) and a background 1006. A dominant motion compensated (DMC) picture 1000 (on the left and also referred to as the decoded reference frame) is first created by forming a GMC morphed picture 1008 rounded or fit within a rectangle 1010 as explained below. A delta correction motion vector (Δmvx, Δmvy) 1016 then may ‘fine tune’ an adjusted (or morphed or warped) position of the foreground object 1012 to a final position 1014. The motion vectors shown herein point from the current frame 1002 to the corresponding position on the reference frame to show where the region, portion, or block comes from as per usual coding diagrams.
While a single delta correction motion vector 1016 of the foreground star shaped object 1014 is shown, in reality there may be at least two delta correction motion vectors at work since one motion vector could be used for the background 1018 (including a zero delta motion vector), and another motion vector used for the foreground (such as the shown delta motion vector). In other alternatives a single motion vector may be used on a block (such as a macroblock or larger), on a tile (such as a 64×64 rectangle or larger), or other partition or portion of a frame that may or may not be formed by grouping blocks, tiles, or other units together.
Referring to
Process 1100 may begin with “obtaining frames of pixel data and having a current frame and a decoded reference frame to use as a motion compensation reference frame for the current frame” 1102.
Thereafter, the process 1100 may comprise “forming a warped global compensated reference frame by displacing at least one portion of the decoded reference frame by using global motion trajectories”. This is explained in detail below. The at least one portion may refer to single portion of the frame, many portions, or the entire frame. The portion may be a block, a tile such as a coding tree block, and/or a region or partition of the frame. The region may or may not be associated with an object in the frame (or in other words, an object shown on the image the frame provides), and may or may not have a boundary that is shaped like the object.
The process 1100 may also comprise “determining a motion vector indicating the motion of the at least one portion and motion from a position based on the warped global compensated reference frame to a position at the current frame”. This may be performed by motion estimation calculations.
The process 1100 may also include “forming a prediction portion based, at least in part, on the motion vectors and corresponding to a portion on the current frame”. Thus, in this case, motion vectors may be applied to adjust the position of the block, tile, region, or object, before the pixel values are used in that portion to form a prediction that may be compared to the corresponding area of an original frame to determine if there is any residual that warrants coding.
Referring to
The process 1200 may be a computer-implemented method for video coding, and comprises “obtaining frames of pixel data and having a current frame and a decoded reference frame to use as a motion compensation reference frame for the current frame” 1202.
The process 1200 also may include “dividing the reference frame into a plurality of portions that are less than the area of the entire frame” 1204. Thus, the frame may be divided into portions that are a uniform unit such as a block or a tile such as a coding tree block, and so forth. Otherwise, the portion may be object based, such as a foreground, a background, a moving object in the frame, or any other object in the frame.
The process 1200 also may include “performing dominant motion compensation comprising applying local global motion compensation on at least one of the portions by displacing the at least one portion of the decoded reference frame by using global motion trajectories at a boundary of the portion” 1206. Specifically, global motion trajectories may be placed at corners of each, or selected, portions on the frame.
The process 1200 also may include “form a prediction portion that corresponds to a portion on the current frame, and by using the pixel values of the displaced portion” 1208. Thus, in this case, the pixel values may be used directly from the warped GMC picture without the use of motion vectors. The local GMC then provides greater accuracy than applying GMC once to the entire frame.
A summary table showing some of the possible options and features described in detail below are listed on the following table for both motion vector options and local global motion compensation options:
Referring to
Process 1300 first may include first obtaining frames of pixel data and having a current frame and a decoded reference frame 1302. As described with encoder 100, a video stream may be provided to an encoder that has a decoding loop 135 in order to find residuals and provide the quantized residuals to a decoder. Thus, frames may be decoded, and used as decoded reference frames to predict yet other frames. A morphing unit such as unit 120 may be used to determine which frames are to be modified or morphed by dominant motion compensation. These frames may already be divided into units such as macroblocks, prediction blocks, and so forth.
Referring to
A GMC morphed reference picture or frame 1500 may be formed 1310 from the GMC morphed picture 1402. This is performed to provide a frame size for ease of computations and comparisons to the original frame. This may include creating a larger padded rectangle 1500 encompassing the GMC morphed picture (a trapezoid) 1402 using the top left coordinate as the reference point (or starting point or connection point). The area outside the quadrilateral 1402 but inside the rectangle 1500 may be filled by padding 1506 which consists of simply copying pixels from the right, top and bottom edges of the quadrilateral 1402, except for corner pixels (areas of overlap) that may be filled by extending both horizontally and vertically and averaging pixels. Areas where the quadrilateral extends out of the rectangle 1500 may be cut or snipped. By one approach, this GMC morphed reference frame is used for motion compensation going forward. It will be understood that the rectangle formed based on the warped picture itself also may be referred to herein as the warped reference frame since it includes warped pixel locations of an image.
By another approach, a virtual GMC morphed picture 1600 (shown in dashed line) may optionally be formed 1312 to proceed with the motion compensation. This may be provided when the system has sufficient compute power to handle division efficiently since such further warping results in significant computation load. Otherwise, the motion compensation may continue with the warped reference rectangle 1500 as explained above.
To provide the virtual GMC morphed picture 1600, the reference picture may be extended to generate virtual reference picture 1600 such that the width and height becomes a power of two. For example if the reference picture 1500 has a width of 720 and height of 480, then the virtual ref picture would have a width of 1024 and height of 512. As before motion trajectories 1602, 1604, and 1606 may be computed for each of the three vertices (with the fourth vertex being unconstrained), and applied to the vertices of the virtual reference picture 1600 (rather than being applied to the reference picture 1400 as done before). The resulting warped quadrilateral (due to application of motion trajectories) is also shown and referred to as the virtual GMC morphed picture 1600. The reason for using virtual GMC morphed picture 1600 instead of GMC morphed picture 1402 for generation of a morphed reference picture 1600 has to do with the fact that the motion compensation process often involves use of much higher precision (often ⅛th pixel precision) than integer pixel, and thus may require interpolation which requires division for scaling. By working with pictures that are powers of 2, scaling related divisions simply become shifts and are much computationally simpler for a decoder.
By a first optional approach, mathematically the affine transform process is described by the following equations that use affine parameters a, b, c, d, e, f to map a set of points (x, y) in a previous picture to a modified set of points (x′, y′).
x
i
′=a·x
i
+b·y
i
+c (10)
y
i
′=d·x
i
+e·y
i
+f (11)
It will be understand for all equations herein that any of (·), (*) or (x) simply refer to multiplication. Equations (10) and (11) effectively modify or morph the reference frame so it can then be used for more efficient motion compensation for a current frame being analyzed. This model is transmitted as three motion trajectories, one for top-left corner of the picture, one for top-right corner of the picture, and one for bottom-left corner of the picture. Affine parameters are calculated (fixed point arithmetic) for a virtual picture which is assumed to be of width and height of nearest power of 2 number which is greater than the coded picture. This removes division operations at the decoder. Formally, assume for 3 vertices (x0, y0), (x1, y1), (x2, y2) corresponding motion trajectories mt0, mt1, and mt2 are given and can be represented as (dx0,dy0), (dx1, dy1), and (dx2, dy2) say in ⅛ pel units. Where:
x0=0,y0=0
x1=W*8,y1=0
x2=0,y2=H*8
where W is the width of a picture and H is the height of the picture. Then, rounding W and H to powers of 2, derive W′ and H′ as follows.
W′=2r:W′>=w,2r-1<W (12)
H′=2s:H′>=H,2s-1<H (13)
The affine parameters A, B, C, D, E, F can then be calculated as follows.
C=dx0 (14)
F=dy0 (15)
A=W′*((x1+dx1)−(x0+dx0))/W (16)
B=W′*((x2+dx2)−(x0+dx0))/W (17)
D=H′*(((y1+dy1)−(y0+dy0))/H) (18)
E=H′*(((y2+dy2)−(y0+dy0))/H) (19)
Other option to calculate the morphed or warped reference frame are provided below.
Process 1300 also may include define frame portions for motion vectors 1304. Here, three options are provided and described below. In one option, motion vectors are provided on a block-by-block basis, and may be defined prediction blocks 1306 such as macroblocks or other prediction or coding units. By another option, the frame may be divided into tiles, which may be blocks of 64×64 pixels or more. For this option the tiles may be grouped into regions associated with an object such as being part of the background or foreground. A motion vector is then determined for each region. By a third option, the regions may be defined directly without any initial division of large tiles or blocks although the boundary to such regions may be defined to fit small blocks (such as 4×4 pixels).
Referring to
By one example, a current frame 1702 with a star shaped object 1704 to be coded may be divided into blocks A2, B2, C2 for coding, and a GMC Morphed Reference frame 1700 may be derived using a past decoded reference frame and GMC motion trajectories that formed the warped quadrilateral 1708. The blocks of pixels A2, B2, C2 from the current frame 1702 are matched during motion estimation to find closest matches at A1, B1, and C1 respectively. The first block match is offset by delta motion vector Δmv1 (1714), the second block match is offset by delta motion vector Δmv2 (1716), and the third block match is offset by delta motion vector Δmv3 (1718) in the warped and padded GMC Reference picture 1700. While only three blocks are shown, it is assumed that the entire picture may be divided into blocks on a block grid, and for each block, a delta motion vector can be computed that provides the best match in the GMC reference picture 1700. Also, while blocks are shown to be of medium size, generally the blocks can be large, small, or each block can be one of a few permitted sizes, and whatever size may be needed to provide the right tradeoff between the reduction in Motion Compensated Prediction error versus the cost of extra delta motion vector information that needs to be coded and transmitted.
Once the blocks are established, motion estimation may be performed 1326 to determine delta motion vectors based on the warped position of the portion of the frame, which is a block in this case. In a practical coding scenario (explained further via
To get higher prediction efficiency, the delta motion vectors may be kept at ¼ or ⅛ th pel precision. To reduce the cost of sending delta motion vectors, the motion vectors identified herein as (Δmv) can be coded efficiently with prediction following the method similar to coding of normal motion vectors. The process 1300 used with blocks as described herein may not be referred to as GMC prediction since it uses delta motion vectors pointed to source blocks at a GMC reference picture 1700. Rather, this is considered a type of dominant motion compensation (DMC), and may be referred to motion vector DMC. Other forms of DMC exist as described below. This difference (between GMC and DMC) here, however, is not minor. It forms an adjustment to the pixel locations that may significantly decrease prediction error over known GMC providing more efficient coding.
Also, the method described herein is simpler than that discussed for warped reference frame 1000 (
Once the delta motion vectors are established, the process 1300 may continue with form prediction 1328 (or specifically a prediction portion or in this case a prediction block) using the portion identified by the motion vector. A simple technique such as bilinear interpolation may be used for generating the necessary DMC prediction block. More sophisticated methods can also be used as follows:
The following is one method for generating a morphed reference (MRef)
1. (Ref Method) Morphed Reference Using Bilinear Interpolation:
A, B, C, D, E, & F are affine parameters calculated from the three motion trajectories transmitted.
x=(A*j+B*i+C<<r)>>r (20)
y=(D*j+E*i+F<<s)>>s (21)
where (j, i) is current pixel location (on the current frame being analyzed), << and >> are left and right bitwise shifts, and (x, y) is the reference pixel coordinate in ⅛th Pel accuracy on the morphed or modified reference frame.
p
y
=y & 0×7 (22)
p
x
=x & 0×7 (23)
y
0
=y>>3 (24)
x
0
=x>>3 (25)
where (x0, y0) is the integer pel location in the Ref Image (reference frame), and px, py is the ⅛th pel phase, & 0×7 refers to bitwise AND with (the binary value of 7 using 8 bits). These represent four corner points used to find a weighted average value for a pixel in the middle of them. Then, the morphed or modified reference is constructed as follows:
MRef[i][j]=((8−px)*(8−py)*Ref[y0][x0]px*(8−py)*Ref[y0][x0+1]+py*(8−px)*Ref[y0+1][x0]py*px*Ref[y0+1][x0+1]+31)>>6 (26)
where MRef is the morphed reference frame, and recited in a different form:
2. Motion Compensated Morphed Reference Prediction using Bilinear Interpolation & MC Filtering:
By another alternative to determine the morphed reference and predictions, motion vectors and variety of block sizes may be factored into the equations as follows. (iMVx, iMVy) is the transmitted motion vector in Sub-Pel Unit (fs) for a block at (j, i) of size (Wb×Hb). A, B, C, D, E, & F are affine parameters calculated from the three Motion trajectories transmitted. Using separable motion compensation (MC) Filters with filter coefficients h[fs][Nt] of norm T, fs is the Sub-Pel Factor (e.g. 2=Half Pel, 4=Quarter Pel, 8=Eighth Pel), where Nt is the number MC Filter Taps, and
i′=i+(iMVy/fs) (28)
j′=j+(iMVx/fs) (29)
p
i
=iMVy & (fs−1) (30)
p
j
=iMvx & (fs−1) (31)
where (j′, i′) is integer motion adjusted current pixel location in a Morphed Reference Image, and pj, pi are the ⅛th pel phases in the Morphed Reference Image. To create an MRef Image then:
x=(A*j+B*i′+C<<r)>>r (32)
y=(D*j′+E*i′+F<<s)>>s (33)
where (x, y) is the reference pixel coordinate in ⅛th Pel accuracy for location (j′, i′)
p
y
=y & 0×7 (34)
p
x
=x & 0×7 (35)
y
0
=y>>3×0=x>>3 (36)
where (x0, y0) is the integer pel location in Ref Image. px, py is the ⅛th pel phase.
MRef[i′][j′]=((8−px)*(8−py)*Ref[y0][x0]+px*(8−py)*Ref[y0][x0+1]+py*(8−px)*Ref[y0+1][x0]+py*px*Ref[y0+1][x0+1]+31)>>6 (37)
tPredh[m][n]=SUMk(h[pj][k]*MRef[i′+m][j′+n+k])/T, (38)
where:
m=[−N
t/2−1,Hb+Nt/2], (39)
n=[0,Wb−1], (40)
k=[−N
t/2−1,Nt/2] (41)
and,
Predji[m][n]=SUMk(h[pj][k]*tPredh[m+k][n])/T, (42)
where
m=[0,Hb−1], (43)
n=[0,Wb−1], (44)
k=(−Nt/2−1,+Nt/2] (45)
MRef is the morphed reference frame, tPredh is the intermediate Horizontal Interpolation, and Predji is the final Motion Compensated Morphed Reference Prediction.
3. Morphed Reference Using Block MC Filtering:
By yet another alternative, A, B, C, D, E, & F are affine parameters calculated from the three Motion trajectories transmitted. Using separable MC Filters with filter coefficients h[fs][Nt] of norm T. fs is the Sub-Pel Factor (e.g. 2=Half Pel, 4=Quarter Pel, 8=Eighth Pel), and where Nt is the number MC Filter Taps
x=(A*j+B*i+C<<r)>>r (52)
y=(D*j+E*i+F<<s)>>s (53)
where (j, i) is every (Ws×Hs) sub-block location in the current image (typically 4×4, 8×4, or 8×8 sub-blocks), x and y are reference pixel coordinates in ⅛th Pel accuracy.
p
y
=y & 0×7 (54)
p
x
=x & 0×7 (55)
y
0
=y>>3 (56)
x
0
=x>>3 (57)
where (x0, y0) is the integer pel location in the reference frame (Ref Image), and px, py is the ⅛th pel phase.
tPredh[m][n]=SUMk(h[px][k]*Ref[y0+m][x0+n+k])/T, (58)
m=[−N
t/2−1,Hs+Nt/2], (59)
n=[0,Ws−1], (60)
k=[−N
t/2−1,+Nt/2] (61)
MRef[i+m][j+n]=SUMk(h[py][k]*tPredh[m+k][n])/T, (62)
m=[0,Hs−1], (63)
n=[0,Ws−1], (64)
k=[−N
t/2−1,+Nt/2] (65)
where MRef is the morphed reference frame; PredH is the intermediate Horizontal Interpolation.
4. Motion Compensated Morphed Reference Prediction Using Single Loop MC Filtering:
By yet a further alternative that factors in motion vectors and variance in block size, (iMVx, iMVy) is the transmitted Motion Vector in Sub-Pel Units (fs) for a block at (j,i) of size (Wb×Hb). A, B, C, D, E, & F are affine parameters calculated from the three motion trajectories transmitted. Using separable MC Filters with filter coefficients h[fs][Nt] of norm T, fs is the Sub-Pel Factor (e.g. 2=Half Pel, 4=Quarter Pel, 8=Eighth Pel), and Nt is the number MC Filter Taps.
i′=(i+u*Hs)*fs+iMVx (72)
j′=(j+v*Ws)*fs+iMVy (73)
where (j, i) is the current block pixel location, (u, v) is the index of every (Ws×Hs) sub-block within given current block of (Wb×Hb), and (Ws×Hs) sub-block is typically 4×4, 8×4, or 8×8. Below, i′,j′ is motion adjusted current pixel location in fs sub-pel accuracy.
x=((A*j′+B*i′+(C*fs)<<r)>>(r+3) (74)
y=((D*j′+E*i′+(F*fs)<<s)>>(s+3) (75)
where x & y are reference pixel coordinates in fs sub-pel accuracy
p
y
=y & (fs−1) (76)
p
x
=x & (fs−1) (77)
y
0
=y/fs (78)
x
0
=x/fs (79)
where y0, x0 is the integer pel location in Ref Image, px, py is the ⅛th pel phase.
tPredh[m][n]=SUMk(h[px][k]*Ref[y0+m][x0+n+k])/T, (80)
m=[−N
t/2−1,Hs+Nt/2], (81)
n=[0,Ws−1], (82)
k=[−N
t/2−1,+Nt/2] (83)
Predji[u*Hs+m][v*Ws+n]=SUMk(h[py][k]*tPredh[m+k][n])/T,
m=[0,Hs−1], (84)
n=[0,Ws−1], (85)
k=[−N
t/2−1,+Nt/2], (86)
v=[0,Wb/Ws−1], (87)
u=[0,Hb/Hs−1] (88)
where tPredh is the intermediate Horizontal Interpolation, and Predji is the final Motion Compensated Morphed Reference Prediction for block of size Wb×Hb at (j, i).
As described above with
Referring now to
As compared to
In this figure GMC/DMC components are separated and shown explicitly as compared to other morphing related components and logic. The main components are Global or Dominant Motion Estimator & Compensated Prediction Picture/Tile/Partition Generator 1804, Dominant Motion Compensation Local/Picture Buffer 1806, and an interpolation subset 1815 of a Motion Compensated Filtering Predictor 1814, whereas the control logic includes global motion trajectory information gmt and delta motion vectors Δmvs. It will be noted that components of other alternative (non-block based) processes for defining frame portions to be used with motion vectors are shown in dashed line and are described later below.
In operation, decoded and filtered picture/s stored in DPR buffer 1802 are input to the Global or Dominant Motion Estimator & Compensated Prediction Picture/Tile/Part. Generator 1804 that, in the present block-based alternative, performs global motion estimation (GME) producing global motion parameters (represented as gmt trajectory) for an entire frame (where the trajectories are located at the corners of the frame) and generating GMC Ref Picture(s) or regions that are stored in Local/Picture Buffer 1806 for Dominant Motion Compensation. Next, the block motion estimator and partitions motion assembler 1812 performs block based motion estimation resulting in motion vectors or delta motion vectors of blocks (or partitions), and these motion vectors are used by motion compensated predictor 1814 (referred to here as Bi-Tree Partitions Char and Motion Compensated Adaptive Precision Filtering Predictor) that generates prediction blocks by sub-pixel interpolation using interpolation unit 1815. The alternative choices of various sub-pixel interpolators are explained above (alternatives 1 to 4). The output prediction blocks (partitions) from motion compensated predictor 1814 are fed to the Prediction Modes and Reference Types Analyzer 1816. Also, intra prediction blocks (or partitions) are input to the Prediction Modes and reference types analyzer 1816 from an intra directional prediction analyzer and generator 1818. On a block (or partition) basis, the Pred Modes & Ref Types Analyzer 1816 determines the best prediction block from various choices (e.g. DMC prediction is one of the many choices available) and outputs it to a differencer (in portion of the circuit not shown here) that generates a prediction error for coding. Further, the entropy coder 1820, also referred to as the Entropy Encoder Morphing and Synthesis Parameters & MVs, encodes GMC/DMC parameters and data, i.e., gmt, and Δmvs. For simplicity other needed information such as modes and reference information that identifies if a block (partition) uses DMC or some other morphed or synthesized prediction type (or intra type) and further, the picture to use as reference is not shown.
As can be noted, the encoder sub-system 1800 can be relatively extended or modified for other video compression techniques. For instance the DMC approach discussed here can also be made to work with by first extending current H.264, and the upcoming HEVC standards. The HEVC standard does not support GMC or any of the morphing or synthesized prediction modes of NGV, but it does support multiple reference based prediction so before DMC based improved delta mvs can be added, GMC would need to be added first to the HEVC standard. Further, NGV codec uses Tiles and Bi-Tree partitioning for motion compensation, but HEVC uses the concept of Coding Tree Blocks (CTB) or Largest Coding Unit (LCU), and quadtree partitioning into Coding Units (CUs) and small codebook of 8, based partitioning into Prediction Units (PUs), that are functionally similar although they have outwardly different processing structures. Mode and ref type information is available in HEVC but would need to be extended to support GMC/DMC extensions.
Referring to
The entropy decoder 1901 (also called Entropy Decoder Morphing and Synthesis Params & MVs) first decodes morphing and synthesis parameters, motion vectors, delta motion vectors (shown), and modes and reference type decisions (not shown). Decoded and filtered picture/s may be stored in DPR buffer 1902 and are input to Global or Dominant Motion Compensated Prediction Picture/Tile/Partition Generator 1904 that uses decoded gmt parameters to generate GMC Reference Picture(s) or regions that are then stored in a Dominant Motion Compensated Local/Picture Buffer 1906. As in the case of the encoder, a motion compensated prediction unit 1912, also called Bi-Tree Partitions Char and Motion Compensated Adaptive Precision Filtering Predictor, is specific to NGV coding, but in general any type of motion compensated predictor could have been used. The set of prediction-interpolation alternatives (1 to 4) described above may be used by an interpolator unit or sub-system 1915 of the prediction unit 1912. Upon decoding, use of the delta motion vectors by the motion compensated prediction unit 1912 results in dominant motion compensated blocks (partitions) that are sent to the Prediction Mode Selector 1914, along with predictions from an intra directional prediction generator 1916, which uses decoded reference type and mode info to output, for each or multiple block (partition), the best prediction block (partition). This is identical, or similar, to the prediction process in the local prediction loop at the encoder (the encoder also uses analysis/estimation process that is not part of the decoder).
Similar to encoder 1800, decoder 1900 also has a simplified view of other morphing components. Thus, Other Morph. Generator & Picture/s Buffer 1908) and parameters (mop), and synthesis components Other Synth Generator & Picture/s Buffer 1910 and parameters (syp) are not elaborated upon here. Further, as with encoder 1800, the portion of decoder 1900 shown can be adapted to work with extended H.264, or HEVC video standards. As mentioned earlier, the HEVC standard does not include morphing and synthesis types of the NGV codec and only shares the commonality of multiple reference frame prediction, so the HEVC standard would have to be extended to support the DMC prediction modes. The NGV and the HEVC standard also differ in partitionings employed. That difference, however, is mainly superficial since at a functional level the processes employed are similar.
Referring to
Three variations of warped reference frames 2000, 2100, and 2200, are provided to show the use of delta motion vectors for DMC on other portions of a frame rather than, or in addition to, a block-by-block basis. A warped reference frame 2000 (
By grouping of tiles based on whether they predominantly include foreground (FG) or Background (BG) in a picture requires only two delta mvs, one for correction of FG, and the other for correction of BG in the GMC Reference frame 2000. Thus, only two delta mvs 2014, 2016, one for FG another for BG, need be coded and transmitted, and a binary mask that distinguishes FG from BG tiles. Since tile are often large, the overhead bits required for sending the mask is often much smaller than that incurred in sending a single delta motion vector with each block or tile. Further, if for certain tiles, another mode instead of DMC is to be used then such tiles do not need to be included in the FG/BG mask. In other words, the position of the tiles at the warped reference frame are still adjusted by the delta motion vector tile-by-tile such as for tile 2010 or 2012, except now, the same delta motion vector is used for each tile in the group or region. Thus, all foreground tiles have the same vector and all the background tiles have the same vector. This may depend on how the computations proceed and very well may be the same as, or stated as, moving the entire region together with the single motion vector. The approach discussed thus far also applies to the case of more than two regions (multi-object segmentation) and requires sending multiple exclusive tile-to-object/region maps instead of a single FG/BG map. Generally, the portions, regions or region-layers may be groups of tiles that have an association to the same object whether that object is a background, a foreground, a moving object, or any other object. It is contemplated that the tiles may be grouped by other common associations such as to size or position on the frame regardless of the image displayed by the frame, and so forth. This applies to any of the groupings discussed herein. While the term tiles from NGV has been used thus far, if DMC is used in conjunction with HEVC, the same principle applies with the term CTB/LCU replacing the term tile.
Referring to
Specifically, current frame 2102 has some tiles 2108 and 2110 that are complete, be it in FG or BG regions, while many tiles in the FG/BG border are either horizontally (2104) or vertically (2106) by bi-tree partition splits allowing improved accuracy with FG/BG approximate region-layers to be constructed. Further, warped reference frame 2100 generated by first estimating global gmt parameters for the entire current picture with regard to a reference frame, and then warping the reference frame using the computed parameters to compute a warped GMC frame 2116, and then padding it with a boundary extension to create the rectangular warped GMC Reference frame 2100. In this GMC Reference frame 2100, two delta motion vectors 2122 and 2124 are then computed, one for approximate FG (full tiles and half tiles that form the approximate foreground region), and the other for approximate BG (remaining full tiles and half tiles that form the background region) such that these delta mvs 2122 and 2124 further allow improved DMC prediction instead of using collocated approximate FG/BG regions (composed of tiles and bi-tree partitioned tiles). In this case, warped and adjusted tile 2126 is used to position current frame tile 2110, while warped and adjusted tile halves 2120 and 2122 respectfully are references for current frame tile halves 2106 and 2108.
Referring to
Referring again to
Both the current picture (being processed) as well as the past decoded reference picture (one of the pictures from DPR Picture Buffer 1802) are input to the Global or Dominant Motion Estimator & Compensated Prediction Picture Generator 1804 so that delta mv DMC parameters can be computed for each of the two approximate regions and by using the DMC Reference Picture. The merge map is used by the motion compensated predictor (also called the Bi-Tree Partitions Char. and Motion Compensated Adaptive Precision (AP) Filtering Predictor) and by the block motion estimator and partitions motion assembler 1812 (also called the 4×4 Block Motion Estimator ¼ & ⅛ pel Accuracy and Partitions Motion Assembler), of which the latter is used to compute delta motion vectors and the former is used to compute actual motion compensated DMC prediction using these delta mvs. The result is a complete DMC Ref Picture. The DMC approximate region based predictions using DMC Ref Pictures are fed to the Pred Modes & Ref Types Analyzer 1816, along with the input from the intra directional predictor 1818, and the process proceeds as described above on a Tile or partition basis where the Pred Modes & Ref Types Analyzer 1816 determines the best prediction from various choices. Further, the entropy coder 1820 encodes DMC data, such as gmt parameters, Δmvs and merge map for this alternative, along with other data such as mvs, mop (morphing parameters) and syp (synthesis parameters) and mode info, and so forth.
Referring to
The Prediction Mode Selector 1914 uses the tiles or partitions based mode information sent by the encoder via the bitstream to use tiles/partitions of either the DMC approximate FG or BG regions from the DMC Reference Picture, or one of the several available morphed predictors, or synthesized predictors, or intra predictors. The resulting prediction is then added back (this portion of decoder is external) to a decoded quantized prediction error decoded at the decoder to reconstruct a final decoded video picture.
Referring to
This variation of delta mv type DMC incurs additional cost in region-layer representation in terms of a differential with regard to a simple block-based (partition-based) merge map, but in return allows for further reduction of DMC prediction error with regard to the previous tile-based technique while requiring the same number (two) of delta mvs. In fact, all three variations block, grouped tile, and whole region here offer different tradeoffs in terms of complexity, overhead, and DMC prediction error reduction. The block-based variation of warped reference frame 1700 (
Referring to
The region bndry map is used by the motion compensated predictor 2414 (also called the Regions Char. and Motion Compensated Adaptive Precision (AP) Filtering Predictor) and by the regions motion estimator 2412 (also called the Regions Motion Estimator ¼ & ⅛ pel Accuracy) of which the latter is used to compute delta motion vectors while the former is used to compute actual motion compensated DMC prediction using these delta mvs. The DMC region based predictions of FG and BG regions are generated using computed motion vectors for offsetting in GMC Reference frames, and resulting predictions are fed to the Prediction Modes & Ref Types Analyzer 2416. The intra data from an intra directional prediction analyzer and generator 2418 as well as morphed and synthesized predictions may be provided to the analyzer 2416. On a local sub-region basis, the Prediction Modes & Ref Types Analyzer 2416 may determine the best prediction from various choices (e.g. DMC prediction is one of the many choices available), and outputs it to a differencer (in a portion of the circuit not shown here) that generates prediction error for coding. Further, the entropy coder 2420 (called the Entropy Encoder Morphing and Synthesis Parameters & MVs) encodes DMC data, such as gmt parameters, delta mvs and region bndry, along with other data such as mvs, mop (morphing parameters) and syp (synthesis parameters) and mode info (not shown). As mentioned earlier, this type of region based DMC makes best sense in context of overall Region Based Video Coder where a picture is divided into region-layers, and for ease of processing each region may be divided into sub-regions.
Referring to
Referring to
Process 2600 first may include obtaining frames of pixel data and having a current frame and a decoded reference frame 2602 as described previously. Process 2600 may then include define frame portions for local global motion compensation 2604. Three alternative ways to divide a frame into portions is provided and generally look similar to the three divisions used with delta motion vectors, except here some key differences exist. Here, the process 2600 may continue with divide the frame into tiles 2606 instead of blocks and where local GMC is applied to each tile. By another alternative, the process 2600 may include divide the frame into tiles 2608 (and/or sub-tiles 2610) and then group the tiles into regions 2612 so that the same local global motion trajectories are applied to each tile in the same region. Otherwise, the process 2600 may include define object-associated regions 2614 so that local GMC will be applied to each region. Each of these options is explained in detail below. Note on both process
Referring to
The process 2600 may then continue with create dominant motion compensated warped portion. To accomplish this, a reference frame 2700 has individual local gmt GMC morphed tiles (or rectangular regions) that may be reconstructed. Thus, the process 2600 may continue with determine local global motion trajectories 2618 which also may be referred to as dominant motion trajectories (dmts). In this variation, each tile (or rectangular region) is allowed to have its own independent set of gmt DMC (or dmts) parameters (each of three vertices of the tile can be provided independent motion with a fourth corner being dependent) instead of a single set of gmt GMC parameters for the entire frame. For example, reference frame 2700 may have a tile 2706 that is a reference for tile 2704 on the current frame 2702. The tile may be displaced from a position on the initial reference frame by global motion trajectories (gmt) 2712a, 2712b, and 2712c. Similarly, a tile 2708 may be displaced by global motion trajectories 2714a, 2714b, and 2714c, while tile 2710 may be displaced by global motion trajectories 2716a, 2716b, and 2716c, such that each tile or region has its own gmt set (or dmts set). As mentioned above, the dmts or local gmt may be obtained by known processes with the affine methods.
Once the tile or region is displaced by the trajectories, then averaging of pixels in overlapping areas of warped neighboring tiles may be performed, followed by filling in holes also using averaging of values from available nearby pixels to form the region rectangle 2622. This is more sophisticated than the rather simplistic boundary extension performed earlier to create a rectangular morphed GMC reference frame. In another alternative, extensions could be used instead to form tile or region rectangles. The resulting filled picture is then the DMC Ref Picture for this type of DMC and can be used for tile (or rectangular region) based DMC motion compensation. As another option, a further virtual region may be formed 2624 as with the virtual reference frame. The process 2600 may then include form portion predictions using the pixels from the warped portions (or tiles or regions) 2626.
While this method may seem to be motion overhead intensive due to the need to send gmt parameters for each tile (or rectangular region), it is important to note that in this type of DMC, (i) no need exists to transmit picture-wide GMC gmt parameters, (ii) gmt parameters are sent in the bitstream only for tiles (rectangular regions) where they are most effective in reducing DMC prediction error; they are not sent for every tile (or rectangular region). If neighboring tiles (regions) gmt parameters are not sent, simple extended padding instead of averaged padding may be employed. (iii) to further reduce the bit cost of gmt parameters per tile, these parameters can be coded differentially with regard to the immediate previous tile (which may reduce the number of bits needed for each trajectory.
Referring again to
Referring to
Referring to
In more detail, a current frame 2802 may be divided into tiles (or rectangular regions). In this variation, a group of tiles (or approximate region-layers) may have their own independent set of gmt, or more specifically dmt, GMC parameters (each of three vertices of each tile in the group of tiles can be provided the same independent motion with fourth vertex of each tile being dependent). In other words, although a set of motion trajectories is still applied to each tile, all (or selected ones) of the tiles in the same region receive the same trajectories. A frame may be divided into two groups of tiles (or approximate region-layers), one group corresponding to the background (BG), and the other group corresponding to foreground (FG). In this scenario, two sets of gmt GMC parameters, one for FG and the other for BG is sent, along with a tile (rectangular region) based FG/BG map (merge map) via the bitstream to the decoder. The operation, however, may not be limited to division of a picture into two groups of tiles (or approximate region layers), and can the frame may be divided into three, four, or more approximate region-layers where each approximate region-layer has one gmt GMC parameter set that is used and sent to the decoder. In any case, for all tiles that are considered as part of the FG group of tiles (approximate region-layer), a single gmt GMC parameter is computed, and likewise for the complimentary tiles that are part of the BG group of tiles (approximate region-layer), a different gmt GMC parameter is computed.
This is illustrated by
To perform DMC motion compensation, a reference picture is read from DPR picture buffers, and the collocated tile corresponding to each tile of an FG region layer is warped at each of the three vertices (the fourth being free) by using FG gmt parameters. Next, the process is repeated for each BG tile of approximate a BG region-layer using the same picture and using BG gmt parameters. Thus the warped reconstruction of each of approximate FG and BG region layers is completed. For the overlapping areas between warped approximate FG and BG region-layers, an averaging process is used to reconstruct the final pixels. Likewise, in an area of holes (an area not covered by either of the approximate FG or BG region-layers), hole filling is employed using averaged prediction from closest neighboring boundary of approximate FG and BG region-layers. The resulting filled picture is then the DMC Reference Picture, and can be used for tile (or rectangular region) based DMC motion compensation.
Further, while the example of two region layer classification into approximate FG/BG regions is provided, the technique can easily be applied to more than two approximate region-layers. Since in this technique there will always be two or more approximate region-layers, instead of the term gmt, the term dmt (dominant motion trajectory) parameters is used as mentioned above
Referring to
Referring to
Referring again to
As before, the merge map information is then input to the Global or Dominant Motion Estimator & Compensated Prediction Pictures/Tiles/Partitions Generator 1804, the other inputs of which include both the current picture (being processed) as well as the past decoded reference picture (one of the pictures from DPR Picture Buffer 1802) so that gmt DMC parameters can be computed for each of the two approximate regions so that the DMC Reference Picture can be generated. The DMC approximate region based predictions using DMC Reference Pictures are fed directly to the Prediction Modes & Ref Types Analyzer 1816 along with other inputs such as intra prediction and morphed and synthesized predictions. On a tile or partition basis, the Pred Modes & Ref Types Analyzer 1816 determines the best prediction from various choices (e.g. DMC prediction is one of the many choices available) and outputs it to a differencer (in a portion of the circuit not shown here) that generates prediction error for coding. Further, the entropy coder 1820 encodes DMC data, such as dmts parameters and the merge map, along with other data such as mvs, mop (morphing parameters) and syp (synthesis parameters) and mode info (not shown). The other components of
Referring again to
Referring to
As with the other local global motion compensation alternatives, the gmt (or dmt) DMC parameters 3120a, 3120b, and 3120c are applied to the vertices of the boundary, and specifically the bounding box region in the reference frame 3108 to form a warped bounding box region 3016 from the unwarped position 3014, and that represents the warped FG region-layer providing a warped head and shoulder (object) position 3118. Similarly, a set of gmt or dmt DMC parameters 3112a, 3112b, and 3112c are computed for the BG region-layer 3124 by using the frame rectangle with FG the region-layer 3116 masked, and the computed gmt DMC parameters 3112a, 3112b, and 3112c are then applied to vertices of the entire frame 3126, resulting in a warped BG region-layer 3110. Since it is possible for the two morphed region-layers 3110 and 3116 to overlap, the area of overlap 3128 can be reconstructed by averaging overlapping pixels from the two regions. Further, since the two warped region-layers 3110 and 3116 can have holes 3130 within the frame 3108, that area is filled by averaged interpolation from neighboring pixels of both regions. Further, as before, any unfilled area close to the frame border is boundary extended as before. In this method, two sets of gmt DMC trajectories (one for FG region-layer and the other for BG region-layer), as well as FG/BG segmentation boundary map is sent via bitstream to the decoder. This particular variation is best used in context of the encoder that already uses region based coding.
Referring again to
As mentioned earlier this type of region based DMC makes best sense in context of overall region based video encoder coder where a frame is divided into region-layers, and for ease of processing, each region is divided into sub-regions. While it is possible for sub-regions to be tiles or blocks, they could also be arbitrary in shape at some precision (say 4×4 block accuracy). This applies whenever sub-region processing is mentioned herein.
Referring again to
Referring to
Referring to
For encoder 3200, by one approach, each or multiple tiles of a frame from DPR buffer 3202 is input to the Dominant Motion Estimator & Compensated Prediction Picture Generator 3220 for computation of an independent set of gmt DMC parameters (each of three vertices of the tile can be provided independent motion with a fourth one being dependent) instead of a single set of gmt GMC parameters for the entire picture. Using individual corresponding gmt GMC parameters, each tile of a previous reference frame (from DPR buffer 3202) is warped to generate individual local gmt GMC morphed tiles such that some warped tiles result in overlapped pixels and others result in holes not covered by any of the tiles. For areas of overlap of tiles, reconstruction is done by averaging of common pixels, and for the areas of holes, averaging from border pixels of neighboring warped tiles is performed. In the case where some tiles are coded by other coding modes, and thus are missing gmt DMC parameters or are at the boundary of the picture, simple boundary extension is performed. The resulting filled picture is then the DMC Reference Picture and is stored in a Dominant Motion Compensated Prediction Local buffer 3222, and can be used for tile based DMC motion compensation. The term dmt (or its plural dmts) may be used to refer to local tile based gmt DMC parameters to differentiate themselves from gmt itself.
With regard to motion vector-based DMC, the same or a different reference frame from DPR buffer 3202 is input to Global Motion Estimator & Compensated Prediction Picture Generator 3204 that performs global motion estimation (GME) producing global motion parameters (represented as gmt trajectories) and generating a GMC reference frame that is stored in Dominant Motion Compensation Prediction Local/Picture Buffer 3206. Next, a block motion estimation and partitions motion assembler 3212 may perform tile based motion estimation resulting in delta motion vectors of tiles that can be used for correction or motion, and that are used by motion compensated predictor 3214 (referred to here as (Bi-Tree Partitions) Char and Motion Compensated Adaptive Precision Filtering Predictor) and that generates prediction tiles by sub-pixel interpolation using the GMC Reference frame.
The output prediction tiles from the delta mv based DMC and local gmt based DMC (as well other morphed blocks/tiles and synthesized blocks/tiles) are fed to the Prediction Modes & Reference Types Analyzer 3216 as well as inputs such as intra predicted blocks/tiles from the intra directional prediction analyzer and generator 3218. On a block or tile basis, the Prediction Modes & Reference Types Analyzer 3216 determines the best prediction block or tile from various choices. For example, the available choices may include DMC prediction as one of the many morphing choices available including one choice that is local gmt DMC based and another choice that is delta motion vector DMC based. The analyzer 3216 outputs the best prediction to a differencer (in portion of circuit not shown here) that generates prediction error for coding. The analyzer 3216 also outputs a map of the DMC mode selection information (dmsi) on a portion (tile or region or other partition) basis that was used in the processing. Further, the entropy coder 3224 (also called Entropy Encoder Morphing and Synthesis Parameters & MVs) encodes GMC/DMC parameters and data, such as the gmt, Δmvs, dmts, and dmsi.
Referring to
Depending on the DMC mode of a tile as carried by the dmsi map, either Global Motion Compensated Prediction Picture Generator 3204 along with block motion compensated predictor 3212 (shown here as (Bi-Tree Partitions) Char. and Motion Compensated Adaptive Precision (AP) Filtering Predictor), or Dominant Motion Compensated Prediction Pictures/Tiles Generator 3218 is deployed to generate the appropriate DMC motion compensation. However, for many tiles or blocks, the DMC mode may not be used since the encoder selects the best mode which may have been one of the other morphed prediction modes, one of the other synthesized prediction modes, or an intra mode.
Referring now to
In the illustrated implementation, video coding system 3500 may include a processing unit such as a graphics processing unit 3520 with logic circuitry 3550, the like, and/or combinations thereof. For example, logic circuitry 3550 may include encoder system 100 of
Process 3400 may begin with “receive input video frames of a video sequence” 3401, where input video frames of a video sequence may be received via encoder 100 for example. This may include a current frame and a past or previous frame (that is to be used as a reference frame), and that is to be used for motion compensation to reconstruct a predicted frame as the current frame.
The process 3400 also comprises “perform prediction and coding partitioning, quantization, and decoding loop to decode a reference frame” 3402. Here, video frames are coded and then decoded in a decoder loop at an encoder in order to provide coded data that can be accurately obtained by a process that is repeated at the decoder.
The process 3400 also comprises “perform frame-wide global motion compensation to form a warped GMC reference frame” 3403. In this case, global motion trajectories are applied at the corner of the frame to warp the frame for delta motion-vector based DMC. The process may be implemented as explained for the other implementations described herein.
The process 3400 also comprises “determine delta motion vector for individual portions of the warped GMC reference” 3404. This may include first defining the portion as blocks, tiles, regions formed by tiles grouped together, or regions without defining tiles first. This may also create a tile merge map or region boundary map as needed. Motion estimation may then be applied to determine the delta motion vector (Amy) for each portion. This may be performed on all blocks or other portions in a frame, or merely selected blocks or selected portions of a frame.
The process 3400 may then continue with “determine motion vector-based prediction for the individual portions” 3405. Thus, a motion predictor such as those described above (such as predictor 123 or 1814) may determine a prediction portion (prediction block, tile, sub-tile, region, sub-region, or other partition) that is provided to a prediction selector. The prediction, if selected, may be used for comparison to the original frame to determine whether a difference or residual exists that warrants coding.
Alternatively, or additionally, the process 3400 may comprise “perform local-global motion compensation on individual portions” 3406. This process also is described above with the other implementations, and the portion once again may be tiles, regions formed by grouping tiles, or regions without tile groupings, and may be object associated regions. Here, the gmt or more accurately the dmt, are applied at the boundary, such as the corners or vertices, of each portion where the compensation is desired. By one form, this is applied to all of the portions in a frame, but need not always be. A tile merge map or region boundary map may also be formed as needed.
The process 3400 may also comprise “determine local GMC based prediction for the individual portions” 3407. Thus, the pixel values of the warped portions are used as predictions, and provided to a prediction selector on a portion (block, tile, sub-tile, region, sub-region, or other partition) basis as described previously.
The process 3400 also comprises “select best prediction for coding of individual portions” 3408. Particularly, the prediction selector compares the different predictions to the original frame and selects the best fit, or may use other criteria. If a difference exists, the difference or residual is placed in the bitstream for transmission to the decoder.
The process 3400 also comprises “transmit motion data and dominant motion compensation parameters to decoder” 3409. The motion data may include residuals and motion vectors, and the dominant motion compensation parameters may include the indicator or map of the selected prediction (dmsi), any tile merge map and/or region bndry map, the dmt and/or gmt trajectories, as well as the delta motion vectors Δmvs, and these may be provided on a block, tile, region or other portion basis as needed. Where the same trajectories or motion vectors are applied to all or multiple portions in a region, the values may only need to be sent once with an explanatory map.
The process 3400 also comprises “receive and decode bitstream” 3410. This may include parsing the bitstream into dominant motion compensation data, motion or other residual data, and image or frame data, and then entropy decoding the data.
The process 3400 also comprises “perform inverse decoding operations to obtain dominant motion compensation parameters with dominant or global motion compensation trajectories, delta motion vectors, prediction selection map, merge map, and/or boundary map” 3411. This will finally reconstruct that values for each type of data.
The process 3400 also comprises “perform dominant (local global) or motion vector based global motion compensation to obtain DMC reference frames or DMC portions” 3412. Thus, the DMC reference frames may be reconstructed, and the same frame or frame portion may be provided with alternative DMC reference frames or portions. This may include applying local global motion trajectories to a boundary to warp individual portions, and alternatively, applying delta motion vectors to portions on a warped reference frame, in order to obtain the pixel values at the resulting prediction portions.
The process 3400 also comprises “provide the predictions to a predictions selector” 3413 where the best prediction for a frame portion is selected and used to form the final frame for display, storage, further encoding, and so forth.
Various components of the systems described herein may be implemented in software, firmware, and/or hardware and/or any combination thereof. For example, various components of system 300 may be provided, at least in part, by hardware of a computing System-on-a-Chip (SoC) such as may be found in a computing system such as, for example, a smart phone. Those skilled in the art may recognize that systems described herein may include additional components that have not been depicted in the corresponding figures. For example, the systems discussed herein may include additional components such as bit stream multiplexer or de-multiplexer modules and the like that have not been depicted in the interest of clarity.
While implementation of the example processes herein may include the undertaking of all operations shown in the order illustrated, the present disclosure is not limited in this regard and, in various examples, implementation of the example processes herein may include the undertaking of only a subset of the operations shown and/or in a different order than illustrated.
Some additional and/or alternative details related to process 1300, 2600, and 3400, and other processes discussed herein may be illustrated in one or more examples of implementations discussed herein and, in particular, with respect to
As illustrated, imaging device(s) 3501, antenna 3502, processing unit(s) 3520, logic circuitry 3550, video encoder 100, video decoder 200, processor(s) 3503, memory store(s) 3504, and/or display device 3505 may be capable of communication with one another. As discussed, although illustrated with both video encoder 100 and video decoder 200, video coding system 3500 may include only video encoder 100 or only video decoder 200 in various examples.
As shown, in some examples, video coding system 3500 may include antenna 3502. Antenna 3502 may be configured to transmit or receive an encoded bitstream of video data, for example. Further, in some examples, video coding system 3500 may include display device 3505. Display device 3505 may be configured to present video data. As shown, in some examples, logic circuitry 3550 may be implemented via processing unit(s) 3520. Processing unit(s) 3520 may include application-specific integrated circuit (ASIC) logic, graphics processor(s), general purpose processor(s), or the like. Video coding system 3500 also may include optional processor(s) 3503, which may similarly include application-specific integrated circuit (ASIC) logic, graphics processor(s), general purpose processor(s), or the like. In some examples, logic circuitry 3550 may be implemented via hardware, video coding dedicated hardware, or the like, and processor(s) 3503 may implemented general purpose software, operating systems, or the like. In addition, memory store(s) 3504 may be any type of memory such as volatile memory (e.g., Static Random Access Memory (SRAM), Dynamic Random Access Memory (DRAM), etc.) or non-volatile memory (e.g., flash memory, etc.), and so forth. In a non-limiting example, memory store(s) 3504 may be implemented by cache memory. In some examples, logic circuitry 3550 may access memory store(s) 3504 (for implementation of an image buffer for example). In other examples, logic circuitry 3550 and/or processing unit(s) 3520 may include memory stores (e.g., cache or the like) for the implementation of an image buffer or the like.
In some examples, video encoder 100 implemented via logic circuitry may include an image buffer (e.g., via either processing unit(s) 3520 or memory store(s) 3504)) and a graphics processing unit (e.g., via processing unit(s) 3520). The graphics processing unit may be communicatively coupled to the image buffer. The graphics processing unit may include video encoder 100 as implemented via logic circuitry 3550 to embody the various modules as discussed with respect to
Video decoder 200 may be implemented in a similar manner as implemented via logic circuitry 3550 to embody the various modules as discussed with respect to decoder 200 of
In some examples, antenna 3502 of video coding system 3500 may be configured to receive an encoded bitstream of video data. As discussed, the encoded bitstream may include data associated with the coding partition (e.g., transform coefficients or quantized transform coefficients, optional indicators (as discussed), and/or data defining the coding partition (e.g., data associated with defining bi-tree partitions or k-d tree partitions using a symbol-run coding or codebook technique or the like)). Video coding system 3500 may also include video decoder 200 coupled to antenna 3502 and configured to decode the encoded bitstream.
In some implementations, the decoder system may include a video decoder configured to decode an encoded bitstream. In some examples, the video decoder may be further configured to receive the bitstream.
In some embodiments, features described herein may be undertaken in response to instructions provided by one or more computer program products. Such program products may include signal bearing media providing instructions that, when executed by, for example, a processor, may provide the functionality described herein. The computer program products may be provided in any form of one or more machine-readable media. Thus, for example, a processor including one or more processor core(s) may undertake one or more features described herein in response to program code and/or instructions or instruction sets conveyed to the processor by one or more machine-readable media. In general, a machine-readable medium may convey software in the form of program code and/or instructions or instruction sets that may cause any of the devices and/or systems described herein to implement at least portions of the features described herein.
In various implementations, system 3600 includes a platform 3602 coupled to a display 3620. Platform 3602 may receive content from a content device such as content services device(s) 3630 or content delivery device(s) 3640 or other similar content sources. A navigation controller 3650 including one or more navigation features may be used to interact with, for example, platform 3602 and/or display 3620. Each of these components is described in greater detail below.
In various implementations, platform 3602 may include any combination of a chipset 3605, processor 3610, memory 3612, antenna 3613, storage 3614, graphics subsystem 3615, applications 3616 and/or radio 3618. Chipset 3605 may provide intercommunication among processor 3610, memory 3612, storage 3614, graphics subsystem 3615, applications 3616 and/or radio 3618. For example, chipset 3605 may include a storage adapter (not depicted) capable of providing intercommunication with storage 3614.
Processor 3610 may be implemented as a Complex Instruction Set Computer (CISC) or Reduced Instruction Set Computer (RISC) processors, x86 instruction set compatible processors, multi-core, or any other microprocessor or central processing unit (CPU). In various implementations, processor 3610 may be dual-core processor(s), dual-core mobile processor(s), and so forth.
Memory 3612 may be implemented as a volatile memory device such as, but not limited to, a Random Access Memory (RAM), Dynamic Random Access Memory (DRAM), or Static RAM (SRAM).
Storage 3614 may be implemented as a non-volatile storage device such as, but not limited to, a magnetic disk drive, optical disk drive, tape drive, an internal storage device, an attached storage device, flash memory, battery backed-up SDRAM (synchronous DRAM), and/or a network accessible storage device. In various implementations, storage 3614 may include technology to increase the storage performance enhanced protection for valuable digital media when multiple hard drives are included, for example.
Graphics subsystem 3615 may perform processing of images such as still or video for display. Graphics subsystem 3615 may be a graphics processing unit (GPU) or a visual processing unit (VPU), for example. An analog or digital interface may be used to communicatively couple graphics subsystem 3615 and display 3620. For example, the interface may be any of a High-Definition Multimedia Interface, DisplayPort, wireless HDMI, and/or wireless HD compliant techniques. Graphics subsystem 3615 may be integrated into processor 3610 or chipset 3605. In some implementations, graphics subsystem 3615 may be a stand-alone device communicatively coupled to chipset 3605.
The graphics and/or video processing techniques described herein may be implemented in various hardware architectures. For example, graphics and/or video functionality may be integrated within a chipset. Alternatively, a discrete graphics and/or video processor may be used. As still another implementation, the graphics and/or video functions may be provided by a general purpose processor, including a multi-core processor. In further embodiments, the functions may be implemented in a consumer electronics device.
Radio 3618 may include one or more radios capable of transmitting and receiving signals using various suitable wireless communications techniques. Such techniques may involve communications across one or more wireless networks. Example wireless networks include (but are not limited to) wireless local area networks (WLANs), wireless personal area networks (WPANs), wireless metropolitan area network (WMANs), cellular networks, and satellite networks. In communicating across such networks, radio 3618 may operate in accordance with one or more applicable standards in any version.
In various implementations, display 3620 may include any television type monitor or display. Display 3620 may include, for example, a computer display screen, touch screen display, video monitor, television-like device, and/or a television. Display 3620 may be digital and/or analog. In various implementations, display 3620 may be a holographic display. Also, display 3620 may be a transparent surface that may receive a visual projection. Such projections may convey various forms of information, images, and/or objects. For example, such projections may be a visual overlay for a mobile augmented reality (MAR) application. Under the control of one or more software applications 3616, platform 3602 may display user interface 3622 on display 3620.
In various implementations, content services device(s) 3630 may be hosted by any national, international and/or independent service and thus accessible to platform 3602 via the Internet, for example. Content services device(s) 3630 may be coupled to platform 3602 and/or to display 3620. Platform 3602 and/or content services device(s) 3630 may be coupled to a network 3660 to communicate (e.g., send and/or receive) media information to and from network 3660. Content delivery device(s) 3640 also may be coupled to platform 3602 and/or to display 3620.
In various implementations, content services device(s) 3630 may include a cable television box, personal computer, network, telephone, Internet enabled devices or appliance capable of delivering digital information and/or content, and any other similar device capable of unidirectionally or bidirectionally communicating content between content providers and platform 3602 and/display 3620, via network 3660 or directly. It will be appreciated that the content may be communicated unidirectionally and/or bidirectionally to and from any one of the components in system 3600 and a content provider via network 3660. Examples of content may include any media information including, for example, video, music, medical and gaming information, and so forth.
Content services device(s) 3630 may receive content such as cable television programming including media information, digital information, and/or other content. Examples of content providers may include any cable or satellite television or radio or Internet content providers. The provided examples are not meant to limit implementations in accordance with the present disclosure in any way.
In various implementations, platform 3602 may receive control signals from navigation controller 3650 having one or more navigation features. The navigation features of controller 3650 may be used to interact with user interface 3622, for example. In various embodiments, navigation controller 3650 may be a pointing device that may be a computer hardware component (specifically, a human interface device) that allows a user to input spatial (e.g., continuous and multi-dimensional) data into a computer. Many systems such as graphical user interfaces (GUI), and televisions and monitors allow the user to control and provide data to the computer or television using physical gestures.
Movements of the navigation features of controller 3650 may be replicated on a display (e.g., display 3620) by movements of a pointer, cursor, focus ring, or other visual indicators displayed on the display. For example, under the control of software applications 3616, the navigation features located on navigation controller 3650 may be mapped to virtual navigation features displayed on user interface 3622. In various embodiments, controller 3650 may not be a separate component but may be integrated into platform 3602 and/or display 3620. The present disclosure, however, is not limited to the elements or in the context shown or described herein.
In various implementations, drivers (not shown) may include technology to enable users to instantly turn on and off platform 3602 like a television with the touch of a button after initial boot-up, when enabled, for example. Program logic may allow platform 3602 to stream content to media adaptors or other content services device(s) 3630 or content delivery device(s) 3640 even when the platform is turned “off.” In addition, chipset 3605 may include hardware and/or software support for 5.1 surround sound audio and/or high definition 7.1 surround sound audio, for example. Drivers may include a graphics driver for integrated graphics platforms. In various embodiments, the graphics driver may comprise a peripheral component interconnect (PCI) Express graphics card.
In various implementations, any one or more of the components shown in system 3600 may be integrated. For example, platform 3602 and content services device(s) 3630 may be integrated, or platform 3602 and content delivery device(s) 3640 may be integrated, or platform 3602, content services device(s) 3630, and content delivery device(s) 3640 may be integrated, for example. In various embodiments, platform 3602 and display 3620 may be an integrated unit. Display 3620 and content service device(s) 3630 may be integrated, or display 3620 and content delivery device(s) 3640 may be integrated, for example. These examples are not meant to limit the present disclosure.
In various embodiments, system 3600 may be implemented as a wireless system, a wired system, or a combination of both. When implemented as a wireless system, system 3600 may include components and interfaces suitable for communicating over a wireless shared media, such as one or more antennas, transmitters, receivers, transceivers, amplifiers, filters, control logic, and so forth. An example of wireless shared media may include portions of a wireless spectrum, such as the RF spectrum and so forth. When implemented as a wired system, system 3600 may include components and interfaces suitable for communicating over wired communications media, such as input/output (I/O) adapters, physical connectors to connect the I/O adapter with a corresponding wired communications medium, a network interface card (NIC), disc controller, video controller, audio controller, and the like. Examples of wired communications media may include a wire, cable, metal leads, printed circuit board (PCB), backplane, switch fabric, semiconductor material, twisted-pair wire, co-axial cable, fiber optics, and so forth.
Platform 3602 may establish one or more logical or physical channels to communicate information. The information may include media information and control information. Media information may refer to any data representing content meant for a user. Examples of content may include, for example, data from a voice conversation, videoconference, streaming video, electronic mail (“email”) message, voice mail message, alphanumeric symbols, graphics, image, video, text and so forth. Data from a voice conversation may be, for example, speech information, silence periods, background noise, comfort noise, tones and so forth. Control information may refer to any data representing commands, instructions or control words meant for an automated system. For example, control information may be used to route media information through a system, or instruct a node to process the media information in a predetermined manner. The implementations, however, are not limited to the elements or in the context shown or described in
As described above, system 3600 may be embodied in varying physical styles or form factors.
As described above, examples of a mobile computing device may include a personal computer (PC), laptop computer, ultra-laptop computer, tablet, touch pad, portable computer, handheld computer, palmtop computer, personal digital assistant (PDA), cellular telephone, combination cellular telephone/PDA, television, smart device (e.g., smart phone, smart tablet or smart television), mobile internet device (MID), messaging device, data communication device, cameras (e.g. point-and-shoot cameras, super-zoom cameras, digital single-lens reflex (DSLR) cameras), and so forth.
Examples of a mobile computing device also may include computers that are arranged to be worn by a person, such as a wrist computer, finger computer, ring computer, eyeglass computer, belt-clip computer, arm-band computer, shoe computers, clothing computers, and other wearable computers. In various embodiments, for example, a mobile computing device may be implemented as a smart phone capable of executing computer applications, as well as voice communications and/or data communications. Although some embodiments may be described with a mobile computing device implemented as a smart phone by way of example, it may be appreciated that other embodiments may be implemented using other wireless mobile computing devices as well. The embodiments are not limited in this context.
As shown in
While implementation of the example processes herein may include the undertaking of all operations shown in the order illustrated, the present disclosure is not limited in this regard and, in various examples, implementation of the example processes herein may include the undertaking of only a subset of the operations shown and/or in a different order than illustrated.
In addition, any one or more of the operations discussed herein may be undertaken in response to instructions provided by one or more computer program products. Such program products may include signal bearing media providing instructions that, when executed by, for example, a processor, may provide the functionality described herein. The computer program products may be provided in any form of one or more machine-readable media. Thus, for example, a processor including one or more processor core(s) may undertake one or more of the operations of the example processes herein in response to program code and/or instructions or instruction sets conveyed to the processor by one or more machine-readable media. In general, a machine-readable medium may convey software in the form of program code and/or instructions or instruction sets that may cause any of the devices and/or systems described herein to implement at least portions of the video systems as discussed herein.
As used in any implementation described herein, the term “module” refers to any combination of software logic, firmware logic and/or hardware logic configured to provide the functionality described herein. The software may be embodied as a software package, code and/or instruction set or instructions, and “hardware”, as used in any implementation described herein, may include, for example, singly or in any combination, hardwired circuitry, programmable circuitry, state machine circuitry, and/or firmware that stores instructions executed by programmable circuitry. The modules may, collectively or individually, be embodied as circuitry that forms part of a larger system, for example, an integrated circuit (IC), system on-chip (SoC), and so forth. For example, a module may be embodied in logic circuitry for the implementation via software, firmware, or hardware of the coding systems discussed herein.
Various embodiments may be implemented using hardware elements, software elements, or a combination of both. Examples of hardware elements may include processors, microprocessors, circuits, circuit elements (e.g., transistors, resistors, capacitors, inductors, and so forth), integrated circuits, application specific integrated circuits (ASIC), programmable logic devices (PLD), digital signal processors (DSP), field programmable gate array (FPGA), logic gates, registers, semiconductor device, chips, microchips, chip sets, and so forth. Examples of software may include software components, programs, applications, computer programs, application programs, system programs, machine programs, operating system software, middleware, firmware, software modules, routines, subroutines, functions, methods, procedures, software interfaces, application program interfaces (API), instruction sets, computing code, computer code, code segments, computer code segments, words, values, symbols, or any combination thereof. Determining whether an embodiment is implemented using hardware elements and/or software elements may vary in accordance with any number of factors, such as desired computational rate, power levels, heat tolerances, processing cycle budget, input data rates, output data rates, memory resources, data bus speeds and other design or performance constraints.
One or more aspects of at least one embodiment may be implemented by representative instructions stored on a machine-readable medium which represents various logic within the processor, which when read by a machine causes the machine to fabricate logic to perform the techniques described herein. Such representations, known as “IP cores” may be stored on a tangible, machine readable medium and supplied to various customers or manufacturing facilities to load into the fabrication machines that actually make the logic or processor.
While certain features set forth herein have been described with reference to various implementations, this description is not intended to be construed in a limiting sense. Hence, various modifications of the implementations described herein, as well as other implementations, which are apparent to persons skilled in the art to which the present disclosure pertains are deemed to lie within the spirit and scope of the present disclosure.
The following examples pertain to further implementations.
In one example, a computer-implemented method for video coding comprising: obtaining frames of pixel data and having a current frame and a decoded reference frame to use as a motion compensation reference frame for the current frame; forming a warped global compensated reference frame by displacing at least one portion of the decoded reference frame by using global motion trajectories; determining a motion vector indicating the motion of the at least one portion and motion from a position based on the warped global compensated reference frame to a position at the current frame; and forming a prediction portion based, at least in part, on the motion vectors and corresponding to a portion on the current frame.
By another example, the method also may be comprising wherein the at least one portion is at least one of: (1) a block of pixels used as a unit to divide the current frame and the reference frame into a plurality of the blocks; (2) at least one tile of pixels, each tile being at least 64×64 pixels, and used as a unit to divide the current frame and the reference frame into a plurality of the tiles; the method (2) comprising at least one of: (a) grouping tiles together based on common association with an object in the frame to form the at least one portion; and forming a single motion vector for each group of tiles, (b) grouping the tiles based on a merge map transmittable from an encoder to a decoder; or (3) a region of pixels shaped and sized depending on an object associated with the region, wherein a boundary of the region is at least one of: a shape that resembles the shape of the object associated with the region, and a rectangle placed around the object associated with the region; wherein the region is associated with at least one of: a background of the frame, a foreground of the frame, and a moving object in the frame; the method comprising defining the region based on a boundary map transmittable from an encoder to a decoder; wherein forming a warped global compensated reference frame comprises using the global motion trajectories at the outer corners of the frame; wherein forming a warped global compensated reference frame comprises using an affine or perspective global motion compensation method; wherein the at least one portion comprises a frame divided into a background and a foreground, and wherein determining motion vectors comprises providing the background and foreground each with one motion vector; the method comprising performing dominant motion compensation comprising locally applied global motion compensation so that at least one other set of global motion trajectories are used at corners of at least one region on the frame that is less than the entire frame to form a displaced region; and using the pixel values of the displaced region to form a prediction region that corresponds to a region on the current frame; the method comprising at least one of: (a) performing local global motion compensation on multiple regions of the frame by using a different set of global motion trajectories on each region; (b) wherein each region is a tile, and dividing the frame into the tiles, and wherein each tile has a set of global motion trajectories; (c) providing the option to perform local global motion compensation on a fraction of a tile in addition to entire tiles; wherein each region is shaped and sized depending on an object associated with the region; wherein the object is one of: a foreground, a background, and an object moving in the frame; the method comprising providing the option on the at least one region on a region-by-region basis to select a prediction formed by: (1) a motion vector to form a prediction for the at least one region and using global motion compensation applied to the entire frame, or (2) applying local global motion compensation with a set of global motion trajectories at the region and using displaced pixel values of the region to form a prediction; the method comprising applying local global motion compensation with a set of global motion trajectories applied at a region of the reference frame that has an area less than the entire reference frame, and using motion vectors to form a prediction for the at least one region; the method comprising providing the option to select a mode for a frame among: (1) use the dominant motion compensated reference frame prediction, (2) use blended prediction of multiple dominant motion compensated reference frames, (3) use dominant motion compensated reference with differential translational motion vector for prediction, and (4) use dominant motion compensated reference with differential translational motion vector for prediction, blended with another reference frame; the method comprising at least one of (a) to (c):
(a) performing motion compensated morphed reference prediction using bilinear interpolation and motion compensation (MC) filter to form a morphed reference frame MRef, tPredh as the intermediate horizontal interpolation, and predji as the final motion compensated morphed reference prediction:
and where:
(iMVx, iMVy) is the transmitted motion vector in Sub-Pel Unit (fs) for a block at (j, i) of size (Wb×Hb); A, B, C, D, E, & F are affine parameters calculated from the three Motion trajectories transmitted; using separable motion compensation (MC) Filters with filter coefficients h[fs][Nt] of norm T, fs is the Sub-Pel Factor (e.g. 2=Half Pel, 4=Quarter Pel, 8=Eighth Pel), where Nt is the number MC Filter Taps, and
i′=i+(iMVy/fs).
j′=j+(iMVx/fs)
p
i
=iMVy & (fs−1)
p
j
=iMvx & (fs−1)
(j′, i′) is integer motion adjusted current pixel location in Morphed Reference Image. pj, pi are the ⅛th pel phases in the Morphed Reference Image;
x=(A*j+B*i′+C<<r)>>r
y=(D*j′+E*i′+F<<s)>>s
where (x, y) is the reference pixel coordinate in ⅛th Pel accuracy for location (j′, i′)
p
y
=y & 0×7
p
x
=x & 0×7
y
0
=y>>3
x
0
=x>>3
where (x0, y0) is the integer pel location in Ref Image. px, py is the ⅛th pel phase;
MRef[i′][j′]=((8−px)*(8−py)*Ref[y0][x0]+px*(8−py)*Ref[y0][x0+1]+py*(8−px)*Ref[y0+1][x0]+py*px*Ref[y0+1][x0+1]+31)>>6
tPredh[m][n]=SUMk(h[pj][k]*MRef[i′+m][j′+n+k])/T,
Predji[m][n]=SUMk(h[pj][k]*tPredh[m+k][n])/T,
(b) performing morphed reference prediction using block motion compensation (MC) filtering to form a morphed reference frame Mref, and Predh as the intermediate horizontal interpolation:
and where A, B, C, D, E, & F are affine parameters calculated from the three Motion trajectories transmitted; using separable MC Filters with filter coefficients h[fs][Nt] of norm T; fs is the Sub-Pel Factor (e.g. 2=Half Pel, 4=Quarter Pel, 8=Eighth Pel), and where Nt is the number MC Filter Taps
x=(A*j+B*i+C<<r)>>r
y=(D*j+E*i+F<<s)>>s
(j, i) is every (Ws x sub-block location in current image, x and y are reference pixel coordinates in ⅛th Pel accuracy;
p
y
=y & 0×7
p
x
=x & 0×7
y
0
=y>>3
x
0
=x>>3
(x0, y0) is the integer pel location in the reference frame (Ref Image); px, py is the ⅛th pel phase.
tPredh[m][n]=SUMk(h[px][k]*Ref[y0+m][x0+n+k])/T,
MRef[i+m][j+n]=SUMk(h[py][k]*tPredh[m+k][n])/T,
(c) performing motion compensated morphed reference prediction using single loop motion compensation (MC) filtering to form a morphed reference (Mref) and predictions tPredh as the intermediate horizontal interpolation, and Predji as the final motion compensated morphed reference prediction for block of size Wb×Hb at (j, i):
and where:
(iMVx, iMVy) is the transmitted Motion Vector in Sub-Pel Units (fs) for a block at (j, i) of size (Wb×Hb). A, B, C, D, E, & F are affine parameters calculated from the three Motion trajectories transmitted; using separable MC Filters with filter coefficients h[fs][Nt] of norm T, fs is the Sub-Pel Factor (e.g. 2=Half Pel, 4=Quarter Pel, 8=Eighth Pel), and Nt is the number MC Filter Taps;
i′=(i+u*Hs)*fs+iMVx
j′=(j+v*Ws)*fs+iMVy
where (j, i) is the current block pixel location, (u, v) is the index of every (Ws×Hs) sub-block within given current block of (Wb×Hb), and (Ws×Hs) sub-block. Below, i′, j′ is motion adjusted current pixel location in fs sub-pel accuracy,
x=((A*j′+B*i′+(C*fs)<<r)>>(r+3)
y=((D*j′+E*i′+(F*fs)<<s)>>(s+3)
where x & y are reference pixel coordinates in fs sub-pel accuracy
p
y
=y & (fs−1)
p
x
=x & (fs−1)
y
0
=y/fs
x
0
=x/fs
where y0, x0 is the integer pel location in Ref Image, px, py is the ⅛th pel phase;
tPredh[m][n]=SUMk(h[px][k]*Ref[y0+m][x0+n+k])/T,
m=[−N
t/2−1,Hs+Nt/2],
n=[0,Ws−1],
k=[−N
t/2−1,+Nt/2]
Predji[u*Hs+m][v*Ws+n]=SUMk(h[py][k]*tPredh[m+k][n])/T,
m=[0,Hs−1],
n=[0,Ws−1],
k=[−N
t/2−1,+Nt/2],
v=[0,Wb/Ws−1],
u=[0,Hb/Hs−1].
By another approach, a computer-implemented method for video coding, comprising: obtaining frames of pixel data and having a current frame and a decoded reference frame to use as a motion compensation reference frame for the current frame; dividing the reference frame into a plurality of portions that are less than the area of the entire frame; performing dominant motion compensation comprising applying local global motion compensation on at least one of the portions by displacing the at least one portion of the decoded reference frame by using global motion trajectories at a boundary of the portion; and forming a prediction portion that corresponds to a portion on the current frame, and by using the pixel values of the displaced portion.
By yet another approach, the method may also be comprising performing local global motion compensation on a plurality of the portions by using a different set of global motion trajectories on each portion of the plurality of portions; wherein each portion is a tile, the method comprising dividing the frame into the tiles, and wherein each tile has a set of global motion trajectories; the method comprising providing the option to perform local global motion compensation on a fraction of a tile in addition to entire tiles; wherein local global motion compensation trajectories are provided to half-tiles or quarter-tiles; the method comprising at least one of: (a) grouping a plurality of the tiles into a region, and applying the same global motion trajectories on the tiles within the same region, and different sets of global motion trajectories depending on the region, and (b) grouping a plurality of the portions into a region, and applying the same global motion trajectories on the portions within the same region, and different sets of global motion trajectories depending on the region; wherein each portion is shaped and sized depending on an object associated with the portion; wherein the object is one of: a foreground, a background, and an object moving in the frame; wherein the portion is a rectangle placed about the object; the method comprising forming a portion of the background of the reference frame, and a portion of the foreground of the reference frame each with a different set of local global motion trajectories for each portion.
By a further example, a coder comprising: an image buffer; and a graphics processing unit configured to: obtain frames of pixel data and having a current frame and a decoded reference frame to use as a motion compensation reference frame for the current frame; divide the reference frame into a plurality of portions that are less than the area of the entire frame; perform dominant motion compensation comprising applying local global motion compensation on at least one of the portions by displacing the at least one portion of the decoded reference frame by using global motion trajectories at a boundary of the portion; and form a prediction portion that corresponds to a portion on the current frame and by using the pixel values of the displaced portion.
By a further example, the coder may have the graphics processing unit configured to: perform local global motion compensation on a plurality of the portions by using a different set of global motion trajectories on each portion of the plurality of portions; wherein each portion is a tile, the graphics processing unit configured to divide the frame into the tiles, and wherein each tile has a set of global motion trajectories; the graphics processing unit configured to provide the option to perform local global motion compensation on a fraction of a tile in addition to entire tiles; wherein local global motion compensation trajectories are provided to half-tiles or quarter-tiles; the graphics processing unit configured to at least one of: (a) group a plurality of the tiles into a region, and apply the same global motion trajectories on the tiles within the same region, and different sets of global motion trajectories depending on the region; and (b) group a plurality of the portions into a region, and apply the same global motion trajectories on the portions within the same region, and different sets of global motion trajectories depending on the region; wherein each portion is shaped and sized depending on an object associated with the portion; wherein the object is one of: a foreground, a background, and an object moving in the frame; wherein the portion is a rectangle placed about the object; the graphics processing unit configured to form a portion of the background of the reference frame, and a portion of the foreground of the reference frame each with a different set of local global motion trajectories for each portion.
By yet another approach, a coder may comprise: an image buffer; and a graphics processing unit configured to: obtain frames of pixel data and having a current frame and a decoded reference frame to use as a motion compensation reference frame for the current frame; form a warped global compensated reference frame by displacing at least one portion of the decoded reference frame by using global motion trajectories; determine a motion vector indicating the motion of the at least one portion and motion from a position based on the warped global compensated reference frame to a position at the current frame; and form a prediction portion based, at least in part, on the motion vectors and corresponding to a portion on the current frame.
By yet a further approach, the coder may comprise wherein the at least one portion is at least one of: (1) a block of pixels used as a unit to divide the current frame and the reference frame into a plurality of the blocks; (2) at least one tile of pixels, each tile being at least 64×64 pixels, and used as a unit to divide the current frame and the reference frame into a plurality of the tiles; the graphics processing unit of (2) being configured to at least one of: (a) group tiles together based on common association with an object in the frame to form the at least one portion; and form a single motion vector for each group of tiles, (b) group the tiles based on a merge map transmittable from an encoder to a decoder; (3) a region of pixels shaped and sized depending on an object associated with the region, wherein a boundary of the region of (3) is at least one of: a shape that resembles the shape of the object associated with the region, and a rectangle placed around the object associated with the region; wherein the region is associated with at least one of: a background of the frame, a foreground of the frame, and a moving object in the frame; the graphics processing unit being configured to define the region based on a boundary map transmittable from an encoder to a decoder; wherein form a warped global compensated reference frame comprises using the global motion trajectories at the outer corners of the frame; wherein form a warped global compensated reference frame comprises using an affine or perspective global motion compensation method. The coder wherein the at least one portion comprises a frame divided into a background and a foreground, and wherein determining motion vectors comprises providing the background and foreground each with one motion vector; the graphics processing unit configured to perform dominant motion compensation comprising locally applied global motion compensation so that at least one other set of global motion trajectories are used at corners of at least one region on the frame that is less than the entire frame to form a displaced region; and use the pixel values of the displaced region to form a prediction region that corresponds to a region on the current frame; the graphics processing unit configured to at least one of: perform local global motion compensation on multiple regions of the frame by using a different set of global motion trajectories on each region; wherein each region is a tile, and dividing the frame into the tiles, and wherein each tile has a set of global motion trajectories; provide the option to perform local global motion compensation on a fraction of a tile in addition to entire tiles; wherein each region is shaped and sized depending on an object associated with the region; wherein the object is one of: a foreground, a background, and an object moving in the frame; the graphics processing unit being configured to provide the option on the at least one region on a region-by-region basis to select a prediction formed by: (1) a motion vector to form a prediction for the at least one region and using global motion compensation applied to the entire frame, or (2) apply local global motion compensation with a set of global motion trajectories at the region and using displaced pixel values of the region to form a prediction; the graphics processing unit configured to apply local global motion compensation with a set of global motion trajectories applied at a region of the reference frame that has an area less than the entire reference frame, and use motion vectors to form a prediction for the at least one region; the graphics processing unit configured to provide the option to select a mode for a frame among: (1) use the dominant motion compensated reference frame prediction, (2) use blended prediction of multiple dominant motion compensated reference frames, (3) use dominant motion compensated reference with differential translational motion vector for prediction, and (4) use dominant motion compensated reference with differential translational motion vector for prediction, blended with another reference frame; the graphics processing unit configured to at least one of (a) to (c):
(a) perform motion compensated morphed reference prediction using bilinear interpolation and motion compensation (MC) filter to form a morphed reference frame MRef, tPredh as the intermediate horizontal interpolation, and predji as the final motion compensated morphed reference prediction:
and where:
(iMVx, iMVy) is the transmitted motion vector in Sub-Pel Unit (fs) for a block at (j, i) of size (Wb×Hb); A, B, C, D, E, & F are affine parameters calculated from the three Motion trajectories transmitted; using separable motion compensation (MC) Filters with filter coefficients h[fs][Nt] of norm T, fs is the Sub-Pel Factor (e.g. 2=Half Pel, 4=Quarter Pel, 8=Eighth Pel), where Nt is the number MC Filter Taps, and
i′=i+(iMVy/fs).
j′=j+(iMVx/fs)
p
i
=iMVy & (fs−1)
p
j
=iMvx & (fs−1)
(j′, i′) is integer motion adjusted current pixel location in Morphed Reference Image. pj, pi are the ⅛th pel phases in the Morphed Reference Image;
x=(A*j+B*i′+C<<r)>>r
y=(D*j′+E*i′+F<<s)>>s
where (x, y) is the reference pixel coordinate in ⅛th Pel accuracy for location (j′, i′)
p
y
=y & 0×7
p
x
=x & 0×7
y
0
=y>>3
x
0
=x>>3
where (x0, y0) is the integer pel location in Ref Image. px, py is the ⅛th pel phase;
MRef[i′][j′]=((8−px)*(8−py)*Ref[y0][x0]+px*(8−py)*Ref[y0][x0+1]+py*(8−px)*Ref[y0+1][x0]+py*px*Ref[y0+1][x0+1]+31)>>6
tPredh[m][n]=SUMk(h[pj][k]*MRef[i′+m][j′+n+k])/T,
Predji[m][n]=SUMk(h[pj][k]*tPredh[m+k][n])/T,
(b) perform morphed reference prediction using block motion compensation (MC) filtering to form a morphed reference frame Mref, and Predh as the intermediate horizontal interpolation:
and where A, B, C, D, E, & F are affine parameters calculated from the three Motion trajectories transmitted; using separable MC Filters with filter coefficients h[fs][Nt] of norm T; fs is the Sub-Pel Factor (e.g. 2=Half Pel, 4=Quarter Pel, 8=Eighth Pel), and where Nt is the number MC Filter Taps
x=(A*j+B*i′+C<<r)>>r
y=(D*j′+E*i′+F<<s)>>s
(j, i) is every (Ws×Hs) sub-block location in current image, x and y are reference pixel coordinates in ⅛th Pel accuracy;
p
y
=y & 0×7
p
x
=x & 0×7
y
0
=y>>3
x
0
=x>>3
(x0, y0) is the integer pel location in the reference frame (Ref Image); px, py is the ⅛th pel phase.
tPredh[m][n]=SUMk(h[px][k]*Ref[y0+m][x0+n+k])/T,
MRef[i+m][j+n]=SUMk(h[py][k]*tPredh[m+k][n])/T,
(c) perform motion compensated morphed reference prediction using single loop motion compensation (MC) filtering to form a morphed reference (Mref) and predictions tPredh as the intermediate horizontal interpolation, and Predji as the final motion compensated morphed reference prediction for block of size Wb×Hb at (j, i):
and where:
(iMVx, iMVy) is the transmitted Motion Vector in Sub-Pel Units (fs) for a block at (j, i) of size (Wb×Hb). A, B, C, D, E, & F are affine parameters calculated from the three Motion trajectories transmitted; using separable MC Filters with filter coefficients h[fs][Nt] of norm T, fs is the Sub-Pel Factor (e.g. 2=Half Pel, 4=Quarter Pel, 8=Eighth Pel), and Nt is the number MC Filter Taps;
i′=(i+u*Hs)*fs+iMVx
j′=(j+v*Ws)*fs+iMVy
where (j, i) is the current block pixel location, (u, v) is the index of every (Ws×Hs) sub-block within given current block of (Wb×Hb), and (Ws×Hs) sub-block. Below, i′, j′ is motion adjusted current pixel location in fs sub-pel accuracy,
x=((A*j′+B*i′+(C*fs)<<r)>>(r+3)
y=((D*j′+E*i′+(F*fs)<<s)>>(s+3)
where x & y are reference pixel coordinates in fs sub-pel accuracy
p
y
=y & (fs−1)
p
x
=x & (fs−1)
y
0
=y/fs
x
0
=x/fs
where y0, x0 is the integer pel location in Ref Image, px, py is the ⅛th pel phase;
tPredh[m][n]=SUMk(h[px][k]*Ref[y0+m][x0+n+k])/T,
m=[−N
t/2−1,HsNt/2],
n=[0,Ws−1],
k=[−N
t/2−1,+Nt/2],
Predji[u*Hs+m][v*Ws+n]=SUMk(h[py][k]*tPredh[m+k][n])/T,
m=[0,Hs−1],
n=[0,Ws−1],
k=[−N
t/2−1,+Nt/2],
v=[0,Wb/Ws−1],
u=[0,Hb/Hs−1].
By one implementation, at least one computer readable memory comprising instructions, that when executed by a computing device, cause the computing device to: obtain frames of pixel data and having a current frame and a decoded reference frame to use as a motion compensation reference frame for the current frame; divide the reference frame into a plurality of portions that are less than the area of the entire frame; perform dominant motion compensation comprising applying local global motion compensation on at least one of the portions by displacing the at least one portion of the decoded reference frame by using global motion trajectories at a boundary of the portion; and form a prediction portion that corresponds to a portion on the current frame and by using the pixel values of the displaced portion.
By another implementation, the computer readable memory may also include wherein the instructions cause the computing device to: perform local global motion compensation on a plurality of the portions by using a different set of global motion trajectories on each portion of the plurality of portions; wherein each portion is a tile, the instructions cause the computing device to divide the frame into the tiles, and wherein each tile has a set of global motion trajectories; the instructions cause the computing device to provide the option to perform local global motion compensation on a fraction of a tile in addition to entire tiles; wherein local global motion compensation trajectories are provided to half-tiles or quarter-tiles; the instructions cause the computing device to at least one of: (a) group a plurality of the tiles into a region, and apply the same global motion trajectories on the tiles within the same region, and different sets of global motion trajectories depending on the region; and (b) group a plurality of the portions into a region, and apply the same global motion trajectories on the portions within the same region, and different sets of global motion trajectories depending on the region; wherein each portion is shaped and sized depending on an object associated with the portion; wherein the object is one of: a foreground, a background, and an object moving in the frame; wherein the portion is a rectangle placed about the object; the instructions cause the computing device to form a portion of the background of the reference frame, and a portion of the foreground of the reference frame each with a different set of local global motion trajectories for each portion.
By a further example, at least one computer readable memory comprising instructions, that when executed by a computing device, cause the computing device to: obtain frames of pixel data and having a current frame and a decoded reference frame to use as a motion compensation reference frame for the current frame; form a warped global compensated reference frame by displacing at least one portion of the decoded reference frame by using global motion trajectories; determine a motion vector indicating the motion of the at least one portion and motion from a position based on the warped global compensated reference frame to a position at the current frame; and form a prediction portion based, at least in part, on the motion vectors and corresponding to a portion on the current frame.
By yet a further example, the computer readable memory may also comprise wherein the at least one portion is at least one of: (1) a block of pixels used as a unit to divide the current frame and the reference frame into a plurality of the blocks; (2) at least one tile of pixels, each tile being at least 64×64 pixels, and used as a unit to divide the current frame and the reference frame into a plurality of the tiles; the instructions causing the computing device of (2) to at least one of: (a) group tiles together based on common association with an object in the frame to form the at least one portion; and forming a single motion vector for each group of tiles, (b) group the tiles based on a merge map transmittable from an encoder to a decoder. (3) a region of pixels shaped and sized depending on an object associated with the region, wherein a boundary of the region is at least one of: a shape that resembles the shape of the object associated with the region, and a rectangle placed around the object associated with the region; wherein the region is associated with at least one of: a background of the frame, a foreground of the frame, and a moving object in the frame; the instructions causing the computing device to define the region based on a boundary map transmittable from an encoder to a decoder; wherein form a warped global compensated reference frame comprises using the global motion trajectories at the outer corners of the frame; wherein form a warped global compensated reference frame comprises using an affine or perspective global motion compensation method. The memory wherein the at least one portion comprises a frame divided into a background and a foreground, and wherein determining motion vectors comprises providing the background and foreground each with one motion vector; the instructions causing the computing device to perform dominant motion compensation comprising locally applied global motion compensation so that at least one other set of global motion trajectories are used at corners of at least one region on the frame that is less than the entire frame to form a displaced region; and use the pixel values of the displaced region to form a prediction region that corresponds to a region on the current frame; the instructions causing the computing device to at least one of: (a) perform local global motion compensation on multiple regions of the frame by using a different set of global motion trajectories on each region; (b) wherein each region is a tile, and dividing the frame into the tiles, and wherein each tile has a set of global motion trajectories; (c) provide the option to perform local global motion compensation on a fraction of a tile in addition to entire tiles; wherein each region is shaped and sized depending on an object associated with the region; wherein the object is one of: a foreground, a background, and an object moving in the frame; the instructions causing the computing device to provide the option on the at least one region on a region-by-region basis to select a prediction formed by: (1) a motion vector to form a prediction for the at least one region and use global motion compensation applied to the entire frame, or (2) apply local global motion compensation with a set of global motion trajectories at the region and using displaced pixel values of the region to form a prediction; the instructions causing the computing device to apply local global motion compensation with a set of global motion trajectories applied at a region of the reference frame that has an area less than the entire reference frame, and use motion vectors to form a prediction for the at least one region; and the instructions causing the computing device to provide the option to select a mode for a frame among: (1) use the dominant motion compensated reference frame prediction, (2) use blended prediction of multiple dominant motion compensated reference frames, (3) use dominant motion compensated reference with differential translational motion vector for prediction, and (4) use dominant motion compensated reference with differential translational motion vector for prediction, blended with another reference frame.
In a further example, at least one machine readable medium may include a plurality of instructions that in response to being executed on a computing device, causes the computing device to perform the method according to any one of the above examples.
In a still further example, an apparatus may include means for performing the methods according to any one of the above examples.
The above examples may include specific combination of features. However, such the above examples are not limited in this regard and, in various implementations, the above examples may include the undertaking only a subset of such features, undertaking a different order of such features, undertaking a different combination of such features, and/or undertaking additional features than those features explicitly listed. For example, all features described with respect to the example methods may be implemented with respect to the example apparatus, the example systems, and/or the example articles, and vice versa.
| Number | Date | Country | Kind |
|---|---|---|---|
| PCT/US2013/078114 | Dec 2013 | US | national |
The present application claims the benefit of the international application no. PCT/US2013/078114, filed 27 Dec. 2013, the disclosure of which is expressly incorporated herein in its entirety for all purposes.
| Filing Document | Filing Date | Country | Kind |
|---|---|---|---|
| PCT/US2014/024694 | 3/12/2014 | WO | 00 |
