Patent Application
20240364894
Embodiments of the invention relate to video coding, and more specifically to the segmentation of a coded picture into one or more columns and slices.
Digital video capabilities can be incorporated into a wide range of devices, including digital televisions, digital direct broadcast systems, wireless broadcast systems, personal digital assistants (PDAs), laptop or desktop computers, video cameras, digital recording devices, video gaming devices, video game consoles, cellular or satellite radio telephones, and the like. Digital video devices may implement video compression techniques, such as those described in standards like MPEG-2, MPEG-4, both available from the International Organization for Standardization (“ISO”) 1, ch. de la Voie-Creuse, Case postale 56, CH-1211 Geneva 20, Switzerland, or www.iso.org,or ITU-T H.264/MPEG-4, Part 10, Advanced Video Coding (“AVC”), available from the International Telecommunication Union (“ITU”), Place de Nations, CH-1211 Geneva 20, Switzerland or www.itu.int,each of which is incorporated herein by reference in their entirety, or according to other standard or non-standard specifications, to encode and/or decode digital video information efficiently.
A video encoder can receive uncoded video information for processing in any suitable format, which may be a digital format conforming to ITU-R BT 601 (available from the International Telecommunications Union, Place des Nations, 1211 Geneva 20, Switzerland, www.itu.int,and which is incorporated herein by reference in its entirety) or in some other digital format. The uncoded video may be organized both spatially into pixel values arranged in one or more two-dimensional matrices as well as temporally into a series of uncoded pictures, with each uncoded picture comprising one or more of the above-mentioned two-dimensional matrices of pixel values. Further, each pixel may comprise a number of separate components used to represent color in digital format. One common format for uncoded video that is input to a video encoder has, for each group of four pixels, four luminance samples which contain information regarding the brightness/lightness or darkness of the pixels, and two chrominance samples which contain color information (e.g., YCrCb 4:2:0).
One function of video encoders is to translate (more generally “transform”) uncoded pictures into a bitstream, packet stream, NAL unit stream, or other suitable transmission format (all referred to as “bitstream” henceforth), with goals such as reducing the amount of redundancy encoded into the bitstream to thereby increase transmission rates, increasing the resilience of the bitstream to suppress bit errors or packet erasures that may occur during transmission (collectively known as “error resilience”), or other application-specific goals. Embodiments of the present invention provided for at least one of the removal or reduction of redundancy, the increase in error resilience, and implementability of video encoders and/or associated decoders in parallel processing architectures.
One function of video decoders is to receive as its input a coded video in the form of a bitstream that may have been produced by a video encoder conforming to the same video compression standard. The video encoder then translates (more generally “transforms”) the received coded bitstream into uncoded video information that may be displayed, stored, or otherwise handled.
Both video encoders and video decoders may be implemented using hardware and/or software configuration, including combinations of both hardware and software. Implementations of either or both may include the use of programmable hardware components such as general purpose central processing units CPUs, such as those found in personal computers (PCs), embedded processors, graphic card processors, digital signal processors (DSPs), field programmable gate arrays (FPGAs), or others. To implement at least parts of the video encoding or decoding, instructions may be needed, and those instructions may be stored and distributed using one or more non-transitory computer readable media. Computer readable media choices include compact disc read-only memory (CD-ROM), digital videodisc read-only memory (DVD-ROM), memory stick, embedded ROM, or others.
In the following, certain systems, methods and/or aspects relating in at least one broad aspect to video compression and decompression, i.e., the operations performed in a video encoder and/or decoder, will be described. A video decoder may perform all, or a subset of, the inverse operations of the encoding operations. Unless otherwise noted, techniques of video encoding described herein are intended also to encompass the inverse of the described video encoding techniques (namely associated video decoding).
The uncompressed, digital representation of video can be viewed as a sample stream, wherein the samples can be processed by the video display in scan order. One type of boundary often occurring in this sample stream is the boundary between pictures in the sample stream. Many video compression standards recognize this boundary and often divide the coded bitstream at these boundaries, for example through the insertion of a picture header or other metadata at the beginning of each uncoded picture.
For some applications, it may be advantageous to segment the coded picture into smaller data blocks, which segmenting can occur prior to, or during, the encoding. Two use cases for which picture segmentation may be advantageous are described below.
The first such use case involves parallel processing. Previously, standard definition video (e.g., 720×480 or 720×576 pixels) was the largest format in widespread commercial use. More recently HD (up to 1920×1080 pixels) formats as well as 4k (4096×2048 pixels), 8k (8192×4096 pixels), and still larger formats are emerging and finding use in a variety of application spaces. Despite the increase in affordable computing power over the years, as a result of the very large picture sizes associated with some of these newer and larger formats, it is often advantageous to leverage the efficiency of parallel processing to the encoding and decoding processes. Parallel encoding and decoding may occur at the instruction level (e.g., using SIMD), in a pipeline where several video coding units may be processed at different stages simultaneously, or on a large structure basis where collections of video coding sub units are processed by separate computing engines as separate entities (e.g., a multi-core general purpose processor). The last form of parallel processing requires picture segmentation.
The second such use case involves picture segmentation so as to create a bitstream suitable for efficient transport over packet networks. Codecs whose coded video is transported over IP or other packet network protocols can be subject to a maximum transmission unit (“MTU”) size constraint. It is sometimes advantageous for the coded slice size to be such that the resulting packet containing the coded slice is as close to the MTU size as possible without exceeding that size, so as to keep the payload/packetization overhead ratio high, while avoiding fragmentation (and the resulting higher loss probability) by the network.
The MTU size differs widely from network to network. For example, the MTU size of many Internet connections may be set by the smallest MTU size of network infrastructure commonly used on the Internet, which often corresponds to limitations in Ethernet and may be roughly 1500 bytes.
The number of bits in a coded picture depends on many factors such as the source picture's dimensions, the desired quality, the complexity of the content in terms of suitability for prediction, the coding efficiency of the video coding standard, and other factors. However, even at moderate quality settings and content complexity, for sequences of HD resolution and above, the size of an average coded picture easily exceeds the MTU size. For example, a video conferencing encoder can require about 2 Mbits/sec to encode a 720p60 video sequence. This results in an average coded picture size of roughly 33,000 bits or 4125 bytes, which is considerably more than the approximately1500 bytes of the Internet's MTU size. At higher resolutions, the average picture size increases to values significantly above the Internet's MTU size. Assuming a similar compression ratio as in the 720p60 example above, a 4096×2048 (4k) video at 60 fps (4kp60) may require over 300,000 bits, or 25 MTU-sized packets for each coded video picture.
In many video coding standards, a picture segment (or, at least, one form of a picture segment) is known as a “slice”. In the following description, any kind of (e.g., video coding standard based) coded picture fragmentation that breaks any form of in-picture prediction or other coding mechanism may be referred to generally as a “slice”. As such, structures such as the Group Of Blocks (“GOB”) in ITU.T Rec. H.261 or ITU Rec. H.263 (available from the ITU; see above for H.264), slices in H.264 or the MPEG family of standards, may each constitute a “slice” as this term is used herein throughout. However, fragmentation units of RFC3984 or data partitions of H.264 may not constitute a “slice”, as this term is used herein throughout, even if they subdivide the bitstream of a coded picture into smaller datablocks, because they do not break in picture prediction or another coding mechanism.
One advantage of using slices over media unaware segmentation mechanisms, such as, for example, those provided by IP at the routing layer, is that slices are at least to a certain extent independently decodable (as discussed below in more detail). The loss of one slice therefore does not necessarily render the other slices of a coded picture unusable or un-decodable. Depending on the implementation of a fragmentation mechanism, the loss of a fragment, in contrast, may well render many other fragments unusable.
Many or all in-picture prediction mechanisms or coding mechanisms may be? broken by the decoding of a picture header (or equivalent). Whether those prediction mechanisms are broken also by the detection of a slice header may depend on the video compression standard, and the type of slice in use.
In H.264, individual video pictures may be segmented into one or more slices, thereby accommodating applications requiring or otherwise utilized pictures that are partitioned as part of the encoding/decoding process. Slices in H.264 may be independently decodable with respect to motion vector prediction, intra prediction, CA-VLC and CABAC states, and other aspects of the H.264 standard. While this decoding independence may realize increases in error resilience, disallowing the aforementioned prediction across slice boundaries may tend to reduce coding efficiency.
In H.263, a video encoder has more flexibility in selecting which prediction mechanisms are broken through the use of slices or GOBs with non-empty GOB headers. For example, there is a bit included in the picture header, selectable when Annex R is in use, which signals to the decoder that no prediction at all occurs across slice/GOB boundaries. If the bit is not set, though, motion vectors may point outside of the current slice, thereby potentially “importing” sample values that are used for motion compensation within the current slice. Further, loop filtering may incorporate sample values outside of the slice.
In most or all existing video coding standards, with the possible exception of flexible macroblock ordering (“FMO”) used as part of H.264, macroblocks within slices are ordered in raster scan order. Consequently, when video sequences with large picture sizes are partitioned into slices that encompass only a relatively small percentage of all macroblocks in the picture, the slices tend to be elongated when viewed spatially.
The bitstream 106 represents the coded picture corresponding to picture 100, and can include one or more parameter sets 107 as an example of a high level syntax structure, which can include syntax elements relevant to more than one of the coded slices of the picture 100. The parameter set(s) 107 can be followed by one or more slices, each such slice comprising a corresponding slice header 108, 110, and corresponding slice data 109, 111, respectively. Accordingly, in this example, slice header 108 may be associated with slice data 109 and may correspond to slice 102 in matrix 101, while slice header 110 may be associated with slice data 111 and may corresponding to slice 103. The slice headers 108, 110 may include information such as the address of the first macroblock of that respective slice, according to scan order. For example, the second slice 103 when coded into bitstream 106 starts with slice header 110 that includes a first macroblock address of 11′, which designates the address of macroblock 105.
As can be seen in
Also, many entropy coding tools that have two-dimensional properties, such as the coding of motion vectors or intra prediction modes, may be optimized for squared picture aspect ratios. For example, in H.264, the coding of a horizontal motion vector of a given length costs roughly the same number of bits as the coding of a vertical motion vector of the same length. Consequently, these coding tools may yield a better compression for compact slices than for “elongated” slices, such as slices 102 and 103 shown in
Improved coding efficiency for compact slices may further arise from the fact homogenous content, which is more likely to be found in a compact slice, may be more efficiently encoded as compared with the relatively diverse content that is more likely to be found in an elongated slice. As a general though not necessarily absolute rule, picture content is more likely to be homogenous in a compact slice because the spatial distance from the center to the boundaries of the slice is less, on average, for a compact slice than for an elongated slice. Further, having a higher slice area to slice boundary ratio for compact slices means that fewer prediction mechanisms may generally be broken in a given picture, thereby resulting in higher coding efficiency.
In H.264, FMO allows the video encoder to effectively produce rectangular slices by defining rectangular slice groups. FMO is a highly generalized coding tool that was designed to address several issues encountered in video coding. However, from a practical standpoint, FMO tends to be perceived as having a relatively high degree of implementation complexity, resulting in somewhat limited adoption as an aspect of standard video compression. A simpler coding tool that may realize improved coding efficiency, as well as parallel encoding and decoding, may address or ameliorate one or more of the complexity issues associated with a full FMO implementation.
The issue of elongated slices may also appear in an extreme case in many MPEG-2 based encoding schemes. For example, in MPEG-2 encoding, it is often the case that each single row of macroblocks within a picture is encoded into a slice, thereby effectively breaking any in picture prediction mechanisms in the vertical dimension within the picture.
Rectangular slice mode is one of two sub-modes specified in Annex K of H.263, another being “scan order slice mode”, which has properties similar to the slices of H.264 discussed above. Rectangular slices as provided for in H.263 may offer one or more of the earlier described advantages that compact slices provide. However, H.263 requires that the dimensions (specifically the width) of each slice must be conveyed in its corresponding header, which leads to coding inefficiency, for example, in applications in which the slice sizes in the horizontal dimension do not change from picture to picture. In addition, Annex K of H.263 does not specify a minimum slice width that would effectively prevent vertically elongated slices from being used. Vertically elongated slices may introduce implementation difficulties and would not in every case provide the desired coding efficiency advantages that, for the reasons discussed above for horizontally elongated slices, may be provided through use of more compact slices.
Constraining the slice to have a rectangular shape can also be disadvantageous in certain cases. First, rectangular slices may perform sub-optimally in applications for which the bitstreams use transport protocols subject to an MTU. For example, packets may be fragmented if the number of bits within a given packet exceeds the MTU limit imposed on the bitstream, which can be undesirable from at least network performance and error resilience perspectives. Conversely, if the number of bits within a given packet is far below the MTU limit, then the ratio of the number of bits in the transport and slice headers becomes relatively large as compared with the number of bits in the packet payload, thereby leading to coding inefficiencies. Requiring slices to be rectangular in shape limits the encoder's ability to precisely control the number of bits in the coded slice so as to avoid the above-mentioned disadvantages.
Second, rectangular slices may perform sub-optimally in applications that utilize parallel encoding and/or decoding. When encoding and/or decoding in parallel, it is typically advantageous to partition a picture into different parts such that each part of the picture requires approximately the same amount of computational power to encode. By partitioning the picture in this way, each part of the picture may therefore be encoded with nearly the same latency to thereby reduce or minimize lag between the encoding times of different parts of the picture. An encoder constrained to use rectangular slices may not be able to precisely control the amount of CPU capacity required to encode and/or decode each slice and thereby avoid this potential disadvantage.
In order to facilitate parallel decoding of slices belonging to the same coded picture, a decoder will generally assign coded picture segments to the various processors, processor cores, or other independently operating decoding mechanisms made available to the decoder for parallel decoding. Without the use of FMO, this was a generally difficult, in some cases extremely difficult, task for previous video coding standards to handle, as those previous standards would allow too much flexibility in the bit stream generation. For example, in H.264, it is possible that one picture may be coded in a single slice and another picture into dozens of slices within the same bitstream. If parallelization occurs at the slice level, when a picture is coded in a single slice, the processor assigned to decode that picture will need to be provisioned to handle its decoding in full. As a result, without imposing restrictions outside of the video coding standard, there may be comparatively little advantage realized by building parallel decoders if each decoding processor will need to be provisioned to be capable of handling a whole picture in any event.
The slice coding used in many MPEG-2 encoders is widely viewed to be the result of an agreement to utilize an informal Cable Labs specification that suggested a one slice per macroblock row segmentation scheme. Widespread acceptance of this informal specification was eventually gained. While there may have been value in such a segmentation scheme when the first MPEG-2 products became available, around 1995, today the various restrictions associated with the historical specification may significantly limit coding efficiency, although parallelization of decoding of (at least SD-coded) pictures has been a relative non-issue for at least a decade.
A need therefore exists for an improved method and system for picture segmentation that addresses, ameliorates or otherwise provides a useful alternative to the existing shortcomings of video encoders both in terms of MTU size matching and parallel decoding. Accordingly, a solution that addresses, at least in part, the above and other shortcomings is desired.
Embodiments of the invention relate in one or more broad aspects to techniques for segmenting a coded picture into columns and slices.
In some embodiments, a plurality of columns may be defined in one or more syntax elements placed in one or more high level syntax structures, such as slice headers or parameter sets. The columns so defined may be of any width between, for example, one Coded Tree Block (CTB, as defined in, for example in WD4: (B. Bross et. al., “WD4: Working Draft 4 of High-EfficiencyVideo Coding”, available from http://wftp3.itu.int/av-arch/jctvc-site/2011_07_F_Torino/) and the entire horizontal size of the picture as measured in CTBs.
In some embodiments, a column boundary between adjacent columns may break one or more prediction mechanism or loop filtering mechanisms of the video codec. The prediction or loop filter mechanisms broken by column boundaries may be the same or different from those broken by slice boundaries.
In some embodiments, the granularity for calculating and specifying column width can be a Largest Coding Unit (LCU), and a coded tree block (CTB) maybe of the same size as an LCU.
In some embodiments, N columns may be defined by a syntax element N that indicates the number of columns in a picture, and which may be followed by N−1 syntax elements that indicate the corresponding widths of each of the N columns. The width of a designated column, for example the right-most column, may be computed, for example, by subtracting the sum of the widths of the left-most N−1 columns from the width of the picture.
In some embodiments, each column in the picture may have an equal width (except perhaps for one pre-defined column, such as the rightmost column, which can be of lesser width) and a syntax element W may be used to indicate such equal width of every column (except the one pre-defined column. The syntax element W may be placed, for example, in a high level syntax element structure such as a parameter set. In such cases, the width of the one pre-defined column can be determined based on a division operation involving Wand the width of the picture, represented by N, whereby the width of the one pre-defined column may be taken as the remainder when N is divided by W.
In some embodiments, each column in the picture may have an equal width (except perhaps for one pre-defined column, such as the rightmost column, which can be of lesser width, in case the width of a picture in CTBs is not an integer divisible of N.), and a syntax element N may be used to indicate the number of equally wide columns (exclusive perhaps of the one pre-defined column). In such cases, the width of all N columns may be determined by dividing the width of the picture in CTBs by N, with the width of the pre-defined column then equaling the remainder of the division operation.
In some embodiments, a syntax element N may be used to indicate the number of columns in a picture. The equal column width W, in units of LCUs, equals the width of the picture, in units of LCUs, divided by N.
In some embodiments, if the picture width is not an integer divisible by the equal column width W, then one pre-defined column, for example the rightmost column, may have a width that is the remainder of a division operation between the width of the picture (measured, for example, in CTBs) and W−1.
In some embodiments, a slice may encompass one or more CTBs each located within a single column, thereby enabling MTU size matching, while still allowing for parallelization through columns.
In some embodiments, a slice may encompass CTBs that are located in more than one column.
In some embodiments, the width of a column may be adjusted in a slice header or comparable syntax structure, thereby allowing to “load balance” the computational resources in the encoder. Such adjustment of column width may be advantageous if the content complexity is different in different horizontal areas in a source picture sequence.
In some embodiments, a high level syntax element may be used to fix the size of all columns of a picture, thereby allowing an efficient assignment of each column to a single processor or equivalent in a parallel encoding or decoding scenario.
In some embodiments, a constraint for the minimum and/or maximum horizontal width of a column may be imposed. Such constraint may in some cases aid parallelization and can improve coding efficiency. The constraint may, for example, be codified in the level specification of standard, or in a high level syntax element, or can be adapted by optimizing a function that can involve MTU size and average bits per CTB (which can be derived from bitrate, frame rate, and frame size in CTBs).
In some embodiments, a scan order of CTBs may be used for decoding and/or encoding, according to which CTBs may be processed in an ordered sequence, at a first level, by column from left-to-right across all columns in the picture, and a second level below the first level, by CTB from left-to-right and top-to-bottom across all CTBs within a given column.
In one broad aspect, there is provided a method for encoding a video picture comprising a plurality of coded tree blocks (CTBs) arranged into at least two rows and at least one column of CTBs. The method may involve encoding a corresponding column width of each at least one column, each column width greater than zero and a sum of all column widths equal to a horizontal dimension of the video picture, and encoding at least one slice of the video picture. In some cases, the at least one the slice may include at least two CTBs located in at least two rows of CTBs.
In another broad aspect, there is provided one or more non-transitory computer-readable media on which are stored instructions arranged to enable one or more processors to perform a method for video encoding a video picture comprising a plurality of coded tree blocks (CTBs) arranged into at least two rows and at least one column of CTBs. The method may include encoding a corresponding column width of each at least one column, each column width greater than zero and a sum of all column widths equal to a horizontal dimension of the video picture, and encoding at least one slice of the video picture. In some cases, the at least one slice may include at least two CTBs located in at least two rows of CTBs.
In some embodiments, according to either of the above two aspects, the method may further include breaking at least one form of prediction or in-loop filtering at column boundaries between adjacent columns of CTBs.
In some embodiments, according to either of the above two aspects, the method may further include encoding at least one of a number of columns, and a width of a column into a parameter set.
In some embodiments, according to either of the above two aspects, the corresponding column width of at least one column may be variable between at least two rows of CTBs in the video picture. In such embodiments, such variability may be controlled by coding the corresponding column width in a slice header.
In some embodiments, according to either of the above two aspects, the method may further include encoding the plurality of CTBs following a scan order of CTBs, according to which the plurality of CTBs are encoded sequentially by column from left-to-right across a plurality of columns, and sequentially by CTB from left-to-right and top-to-bottom within each of the plurality of columns.
In yet another broad aspect, there is provided a method for decoding a coded video picture comprising a plurality of coded tree blocks (CTBs) arranged into at least two rows and at least one column of CTBs. The method may include obtaining at least one of (i) a corresponding column width of each at least one column and (ii) a number of columns, deriving the corresponding column width of each at least one column (if not obtained from the coded video picture), and decoding at least one slice of the coded video picture. In some cases, the at least one slice may include at least two CTBs located in at least two rows of CTBs.
In yet another broad aspect, there is provided one or more non-transitory computer-readable media on which are stored instructions arranged to enable one or more processors to perform a method for video decoding a coded video picture comprising a plurality of coded tree blocks (CTBs) arranged into at least two rows and at least one column of CTBs. The method may include obtaining at least one of (i) a corresponding column width of each at least one column and (ii) a number of columns, deriving the corresponding column width of each at least one column (if not obtained from the coded video picture), and decoding at least one slice of the coded video picture. In some cases, the at least one slice may include at least two CTBs located in at least two rows of CTBs.
In some embodiments, according to either of the above two aspects, the method may further include breaking at least one form of prediction or in-loop filtering at column boundaries between adjacent columns of CTBs.
In some embodiments, according to either of the above two aspects, at least one of the corresponding column width of the at least one column and the number of columns may be obtained from a parameter set.
In some embodiments, according to either of the above two aspects, the corresponding column width of at least one column may be variable between at least two rows of CTBs in the coded video picture. In such embodiments, such variability may be controlled by coding a column width in a slice header.
In some embodiments, according to either of the above two aspects, the method may further include decoding the plurality of CTBS following a scan order of CTBs, according to which the plurality of CTBs are decoded sequentially by column from left-to-right across a plurality of columns, and sequentially by CTB from left-to-right and top-to-bottom within each of the plurality of columns.
In yet another broad aspect, there is provided a video decoder for a coded video picture comprising a plurality of coded tree blocks (CTBs) arranged into at least two rows and at least one column of CTBs. The video decoder may include a plurality of column decoders, with each column decoder configured to decode slices of the coded video picture located within at least one column of CTBs allocated to that column decoder.
In some embodiments, according to the above aspect, each slice may be fully located within exactly one column.
In some embodiments, according to the above aspect, each column decoder may be a process running on at least one of a core of a multicore processor, or a multiprocessor.
In accordance with further aspects of the described embodiments, there is provided an apparatus such as a data processing system, a method for adapting such apparatus, as well as articles of manufacture such as a non-transitory computer readable medium or product having program instructions recorded thereon for practicing the described method(s).
Various features and advantages of the embodiments of the present invention will become apparent from the following detailed description, taken in combination with the appended drawings, in which:
It will be noted that throughout the appended drawings, like features are identified by like reference numerals.
In the following description, details are set forth to provide an understanding of the invention. In some instances, certain software, circuits, structures and methods have not been described or shown in detail in order not to obscure the invention. The term “data processing system” is used herein to refer to any machine for processing data, including the computer systems, wireless devices, and network arrangements described herein. Embodiments of the present invention may be implemented in any computer programming language provided that the operating system of the data processing system provides the facilities that may support the requirements of these embodiments. Embodiments of the present invention may also be implemented in hardware or in a combination of hardware and software.
Embodiments of the present invention relate to picture segmentation in video compression using columns and slices.
Referring to
Two columns 202, 203 are defined in the picture 201 and are shown divided by a column boundary depicted as boldface, solid line 204. The width of any column can be an integer multiple of the size of a CTB (or an integer multiple of the size of the smallest CTB possible) used within the picture 201. Columns may be of equal or unequal width within the picture 201. As shown, column 202 encompasses CTBs 1 through 8, and column 203 encompasses CTBs 9 through 24. Column 202 is two CTBs wide, whereas column 203 is four CTBs wide. However, the described embodiments are not limited to this number of columns nor to the exemplary column widths shown in
Shown also are two slices defined within picture 201 and delineated by the boldface punctuated bordering line 205. The first slice encompasses the CTBs with CTB addresses 1 through 14. The second slice encompasses CTBs with addresses 15 through 24. CTBs 1-8 of the first slice are located in column 202, whereas CTBs 9-14 of the first slice are located in column 203. Therefore, the first slice covers CTBs that are part of either column 202 and 203. The second slice includes CTBs 15-24, which are all located in column 203. Accordingly, the scan order of CTBs within picture 201 has changed from the per-picture scan order, as shown in
The coded bitstream 206 represents the picture 201 and can contain a parameter set 207 (or can include a reference to a parameter set that has previously been decoded, as multiple pictures can refer to the same parameter set). The parameter set 207 can contain column boundary information 208, identifying directly or indirectly the width of one or more column(s), e.g., columns 202 and 203, as described later. Each of the coded slices can comprise a corresponding slice header 209, 211, and corresponding slice data 210, 212, respectively.
Picture segmentation with slices spanning column boundaries, as shown in
One advantage of allowing slices to span multiple columns across the boundaries between them is to allow for good MTU size matching. With slices able to span multiple columns, none of the slices will be artificially limited in size due to the presence of a column boundary. Another advantage can be that it is possible to represent the whole picture as a single slice (for example, where the MTU size is, in most cases, larger than the coded picture size), while still allowing for parallelization. For convenience, it may be assumed that a column boundary breaks all forms of prediction or in-loop filtering. Under this assumption, one example configuration of a parallel encoder may be as follows. Each processor/core of the parallel decoder may begin encoding the spatial area of a column, starting with the top-left CTB within the column and working sequentially in scan order to the bottom right macroblock, i.e., left-to-right and top-to-bottom. Assuming N columns in a picture and N processors/cores with the parallel decoder, this approach to encoding may results in N sub-bitstreams being generated, one from each processor/core. The N sub-bitstreams may then be combined to form a single bitstream representing the whole picture. A decoder receiving such bitstream may, from the parameter set (e.g., included in the bitstream), ascertain that it is being tasked for parallel decoding in N columns. The decoder may pre-process the combined bitstream by first decoding only syntax elements included in the bitstream (without also performing any computationally intensive reconstruction of CTBs) so as to identify the boundaries between the columns in the bitstream domain. As an example, one such boundary would be located between coded CTB 8 and coded CTB 9 of picture 201. Thereafter, the decoder may extract and/or copy the sub-bitstreams, broken up within the combined bitstream at the pre-identified boundaries, to the processors/cores for processing. The processors/cores within the decoder may independently reconstruct their assigned columns in the pixel domain, which may then be copied together to reconstruct the full image. Other possible implementation strategies are explained in more detail below.
One potential disadvantage of having slices spanning across column boundaries can arise because of difficulties in the implementation of parallelization, for example, when only a (small) subset of prediction and/or loop filtering mechanisms are broken by a column boundary. For example, if column 202 were to be decoded on a first processor, and column 203 were to be decoded on a second processor, then the state of the decoding engine (including information such as CABAC state, motion prediction state, and so on), or at least those parts of it which relate to prediction or loop filter tools not broken by column boundaries, would in at least in some cases have to be handed over from the first processor to the second processor after the decoding of CTB 8, which in column-local scan order is the last macroblock in column 202. The implementation and computational complexity of this handover can depend on, for example, the prediction and loop filter tools that are disrupted by the presence of a column boundary. If, for example, a column boundary breaks any form of prediction but the prediction in the entropy coding, then only information pertaining to the entropy coding engine needs to be handed over. Co-pending U.S. patent application Ser. No. 13/336,475, filed even date herewith, entitled “METHOD AND SYSTEM FOR SELECTIVELYBREAKING PREDICTION IN VIDEO CODING”, which is incorporated herein in its entirety, discloses, among other things, techniques to control the amount of data that needs to be handed over when a slice spans a column boundary.
As seen in
The coded bitstream 309 can contain a parameter set 310 comprising column boundary information 311. While the parameter set 309 may be shown in
The coded picture contains three coded slices (i.e., slice 0, slice 1 and slice 2), each comprising a corresponding slice header 312,314,316, and corresponding slice data 313,315,317, respectively. The coded slice O includes coded slice header 312 and coded slice data 313, which includes the coded information of slice 305, and is depicted having a size (in coded bits) somewhere below the MTU size 318. Coded slice 1 includes coded slice header 314 and coded slice data 315 (corresponding to slice 306). In this example, when coding slice 1, the encoder is able to fill the packet almost up to MTU size 318 with slice header 314 and slice data 315. Note that it may not always be possible to fill the packet up to the last bit of the MTU size 318 because it is not possible to fragment a CTB into more than one slice. Accordingly, subject to this restriction, there may be extra bits leftover after encoding of the last CTB within the slice. Finally, slice 2 includes slice header 316 and slice data 317 (corresponding to slice 307), and is considerably smaller than the MTU size 318 because there are fewer CTBs to be coded in this slice, e.g., as compared to slice 030 and slice 1.
A static width of the columns within a picture can be specified at the sequence level (e.g., in a sequence parameter set), at the picture level (e.g., in a picture parameter set), or in other data structures in the bitstream covering at least one coded picture. Specifying column boundaries once for all slices in a sequence or picture has advantages for applications whose slices do not vary in width from picture to picture, in that the overhead associated with specifying a different slice structure is not required for every picture. It further has the advantage of efficiently allowing the distribution of coded bits among multiple processors or equivalent in a parallel decoding scenario, Dynamic column width adjustments are described later.
Column boundaries may be specified using one or more different data structures, parameters, rules, flags, or other suitable syntax. Using the sequence parameter set (SPS) as the exemplary high level syntax structure to convey column boundaries, the SPS can for example include:
Still other representations of column boundary information which are suitable for different applications may be apparent after familiarization with the present disclosure and are intended to be covered within the scope of the described embodiments.
A video coding standard can advantageously set limits for minimum or maximum column width that can occur in a bitstream. The limits can be specified, for example, in profile and level specifications. For example, a profile designed for high level parallel processing in case of high resolution video decoding, can require that columns must be used with a width no larger than, for example, 1920 samples (a 1080p resolution, which is commonly decodeable in a single core, has a resolution of 1920 horizontal samples). On the other hand, the profile can require, for example, a minimum column width of no less than 640 samples. The lower limit can help to avoid worst-case scenarios and “elongated” slices in the Y dimension. The upper limit can be useful for a number of different reasons. First, as already described, coding efficiency can be improved due to fewer and less elongated slices. Second, the decoder can be based on sub-picture decoders each capable of decoding a single column (henceforth called “column decoders”), but not necessarily capable of handling a full picture. Third, memory requirements in the column decoders can be reduced as line buffers do not need to span the full picture width, but only the column width (plus perhaps implementation dependent padding). Other advantages of such an upper limit on column width may also be apparent.
Described next is the operation of a decoder on a bitstream containing slices and columns, noting how such operation may differ from the decoding process of a bitstream that has been generated without use of columns. For convenience, in the following description it will be assumed that coding of column width was performed according to option (3) above. However, the description should not be understood as being limited only to this type of coding and will apply also to the other two options described above, if/where necessary, by making suitable modification and/or alteration to the description.
In modem video coding standards, parameters pertaining to more than one slice can be stored in higher level syntax structures known as parameter sets. In older standards, these parameters can be stored in headers such as the picture, GOP, or sequence headers. To facilitate the following description, parameter sets are assumed henceforth, although embodiments of the invention may equally utilize other forms of high level syntax structures.
The stored parameters (e.g., “num_column_minus_1” and “columnWidth[ ]”) may be decoded (402) as one part of the parameter set decoding (401). The decoding (402) may be a sub-step of the decoding (401) and may be performed at any appropriate time during the parameter set decoding (401). The exact order may, for example, be constrained by the syntax of the parameter sets themselves (which can mandate the location of the information decoded in 402 relative to other entries in the parameter set).
During the parameter set decoding (401), information relating to the prediction and/or in-loop filtering tools can also be decoded, as described further in co-pending U.S. patent application Ser. No. 13/336,475. Decoding of information relating to prediction and/or in-loop filtering tools is indicated generally in
Following the decoding (401) of parameter set information in method 400, slice decoding and reconstruction can commence.
First, a slice header can be decoded (404). In the slice header, information can be available that allows for the decoder to associate the given slice with the column in which that given slice is included. For example, the slice header can contain a CTB address (i.e., as indicated by “first_ctb_address” in
For example, referring back to
Referring again to
First, the CTBs texture, motion, and other information can be decoded and the current CTB being handled by the decoder can be reconstructed (407), which can involve any or all of the different steps noted above in the context of CTB decoding, while also observing the availability of prediction information outside of the slice or column boundary.
Then, the current CTB address maybe incremented (408). If the incremented CTB address is determined (409) to still be left of the column's right-hand boundary, no further updating of the CTB address is required (indicating in
If, after execution of blocks 408 to 410, the vertical component of the CTB address exceeds the vertical size of the picture, then the slice may span more than one column. In this case, the CTB address may be reset, for example, to the top-left CTB address in the column to the left of the previous column. Decoding and CTB address changes continue in this column.
At this point, it is determined (411) whether more data (indicating one or more CTBs) is available in the slice to be decoded. If it is determined (411) that more data remains in the slice to be decoded, then method 400 proceeds (412) back to block 407, and the decoding continues with the next CTB within the slice. However, ifit is determined (411) that no data remains in the slice to be decoded, then method 400 proceeds (413) back to block 404, and the decoding continues with the next slice in the picture.
According to embodiments of the invention, a decoder can read (initial) column boundary information from high level structures such as picture or sequence parameter sets. For example, this can be implemented in a mechanism that parses first a syntax element (i.e., “num_column_minus1”) from the bitstream. This syntax element may advantageously be part of a picture (or higher) syntax structure, such as a picture parameter set, sequence parameter set, picture header, or comparable structure. If the value of “num_column_minus_1” is greater than 0, then it may be inferred that the number of column boundaries present in a conformant bitstream will be equal to the value stored in this parameter, and there num_column_minus_1 column widths maybe parsed. The widths of the correspondingly parsed columns can be indicated in terms of CTBs. The syntax described above is specified in the style of the H.264 syntax specification, parts of which are summarized below for convenience.
Specifically, in the syntax description provided below, a boldface variable (i.e., lines I and 4 below), in comparison to a plain formatted line, may represent bits that are fetched from the bitstream as a side effect of or otherwise in conjunction with determining the value of the variable. Once read, the value can be used in the decoding process however the variable is no longer set in boldface, since the use of the variable is from that point onwards no longer associated with a bitstream fetch.
num
column
minus
1
If column boundaries can be specified at more than one level, according to the given compression standard, then the syntax of column boundaries can be the same at various levels, for example, sequence and/or picture level. In this case, advantageously, column boundaries indicated at a lower relative level (i.e., the picture level) can take precedence over column boundaries indicated at a higher relative level (i.e., the sequence level).
The parsed columnWidth[ ] values may be used to derive the column boundaries shown below as:
In the following, embodiments of the present invention in which the column boundaries can be adjusted during the coding of a picture, i.e., at the slice level are described.
Referring to
In the following description, for convenience, the slice header is used for conveying modification information because, in certain contexts, the slice header may be the most appropriate syntax structure in H.264 for such purpose. The use of the slice header for information pertaining to more than one slice is not necessarily in line with H.264's syntax model (for example, in H.264, all information that pertains to more than one slice should be included in parameter sets only), but is in line with some older video compression standards (for example, in H.261, GOB header information can be predicted from previous GOB headers, and GOBs are at least in some respects similar to equivalent to H.264 slices). At the same time, if a mechanism were to be included in newer video compression standards, which would allow the coding of properties at the slice level that pertains to more than just the current slice, then such syntax structure could in some cases be a more suitable location for the information presented below than in the slice header itself. For example, included in WD4 is an Adaptation Parameter Set (APS), which is designed to carry information that is likely to change from picture to picture (in contrast to the picture parameter set, which carries information that is unlikely to change from picture to picture). Also being considered in JCT-VC are proposals to allow “updating” an APS at slice boundaries, which, if accepted, would also make the APS an appropriate syntax structure for column width update information, as described below.
In some embodiments of the invention, the column boundaries that have been changed through the override mechanism discussed below revert back to the locations specified at the picture or sequence level when encoding/decoding of the current picture is complete. This feature may, at least partly, compensate for the architectural lack of cleanliness of using the slice header to code information pertaining to more than one slice. For example, one possible reason for disallowing the use of picture headers and similar data structures in H.264 is to avoid error propagation due to corrupted header structures. Reverting back to data conveyed in parameter sets (which have been designed to be very difficult to ever get corrupted) after each picture may ensure that header corruption is limited to one given picture, rather to all pictures following the picture in which the header corruption has occurred.
Shown in
According to the modification 500, column boundary 502 spans the entire vertical height of the picture, while column boundary 501 spans only a portion of the entire vertical height of the picture, i.e., the two uppermost rows of CTBs. The remainder of the column boundary is offset relative to column boundary 501 and, consequently, the leftmost column 503 has variable width. For example, column 503 may be split into two different slices of different width. As shown in
Column boundaries can be modified at the slice level by including boundary offset values to shift a column boundary to, for example, the left, if the offset value is negative, or alternatively to, for example, the right, if the offset value is positive. The slice header of the slice 505 in the leftmost column 503 may, for example, include information to the extent that the original column boundary 501 between the leftmost and middle columns, as defined with respect to slice 504, be shifted by one CTB to the left in respect of the slice 505. As a result, CTB 22 belongs now to the middle column (as opposed to the leftmost column 503), and coding continues in the leftmost column 503 with slice 505, which encompasses CTBs 20, 21, 30, 31, 40, 41, 50, and 51 in this particular example.
There are a number of comer cases that could potentially make implementation of the modification 500 described above seem unnecessarily complex and/or inefficient from a caching viewpoint. However, such comer cases may be effectively handled by a number of restrictions concerning the moving of column boundaries that can be imposed in various embodiments, and which are described further below.
During picture decoding, shifting of the column boundary may come into effect starting at the CTB line in which is locate the address of the first CTB of the slice to contain the boundary offset information in its slice header. Thereafter, the boundary offset may remain in effect for all remaining CTB lines in the picture unless such boundary offset is overridden by another boundary offset in a subsequent slice header.
The boundary offset values may be indicated in units of CTB widths.
The bitstream may contain a “boundary_offset_flag” (i.e., as seen in
boundary_offset_flag
As above, boldface variables (i.e., lines 1, 3, and 4 above), in comparison to a plainformatted line (i.e., line 2), may represent bits that are fetched from the bitstream as a side effect of or otherwise in conjunction with determining the value of the variable.
The method 600 may be initialized (601) by the following settings, which for convenience are expressed in C-style pseudo-code. Such initialization, among other things, copies the parsed boundaries from the variable columnBoundary[ ], which was derived from the high level syntax elements described above, as follows:
Thereafter, it is determined (602) whether the boundary_offset_flag (introduced above) is set. If it is determined (602) that the flag is not set, then updates of column boundaries are not allowed, and the method 600 branches (603) so that the slice is decoded (604) as previously described and specified in the video coding standard to which the docoding complies, for example, which may involve, for each CTB of the slice, processes such as entropy decoding, motion compensation to form a predictor, inverse quantization and inverse transform of coefficients to calculate a residual, adding the residual to the motion compensated predictor, in-CTB deblock-filtering the resulting samples, and so forth. Then if it is determined (605) the last slice in the picture has been decoded (a situation that can be tested by many different mechanisms, such as detecting a change in the value of the picturelD, Temporal Reference, presentation time stamp, or similar syntax element, as well as timeout mechanisms based on RTP timestamps), then the method 600 continues (606) by initializing (601) as described above to process the next picture in the bitstream. However, if it is determined (605) that the slice was not the last one of the picture, then the method 600 continues (607) by determining (602) if the boundary_offset_flag is set.
If it is determined (602) that the boundary_offset_flag is set, then the slice header contains boundary offsets that can be parsed and used to modify colBound[ ], and the method branches (608) for such modification of column boundaries. The modification may commence by an identification of the column to which the slice belongs. Such identification may in turn commence with an initialization (609), for example, which can contain instructions such as the following:
At this point, the method 600 may search through the current state of the column boundaries (e.g., in colBound[ ]) so as to identify the particular column to which the slice being decoded belongs. For example, the method 600 may perform such identification by checking (611) the horizontal address component of the CTB address of the current CTB (which is the first CTB in the slice in question), as determined in the initialization and stored in firstCTBMod, against the stored column boundary values (colBound[ ]). By searching through the column boundaries by incrementing (610) a suitable search index, once a column boundary is identified as being located to the “right” (i.e., as having a larger horizontal address in terms of scan order) of the current slice's first CTB (as expressed in firstCTBMod), the search is finished and the method 600 exits (612) the search loop with the surviving state of the search index i.
The boundaries of the column to which the slice being decoded belongs may be updated (613), according to the surviving state of the search index i, by adding the difference information leftOffset and rightOffset, as initialized (609), to the left and right column boundaries Such updating may be achieved, for example, by the following:
At this point, the method 600 proceeds to the slice decoding (604) as described above.
In some case, it may be advantageous and/or convenient to impose constraints on the column boundary indication including the following constraint:
Such constraint requires columns to be numbered from the left to the right in the picture and also implies that offset values cannot be used to switch column boundary order.
The leftOffset, rightOffset and subsequently derived colBound[ ] values may also be constrained such that each resulting column has a width which is greater than or equal to a minimum column width. For example, the minimum column width may be level specific and/or specified in the video compression standard.
In some cases, the leftOffset and rightOffset may not be used to modify a colBound[ ] used as a slice boundary for a previously coded slice.
In some cases, the value of colBound[0] may not be changed from its initial value of ‘0’ to reflect that the first numbered column begins at the leftmost edge of the picture. Consequently, the value of leftOffset would be ‘0’ in such cases when i=1.
In some cases, the value of colBound[num_columns_minus1+1] may not be changed from its initial value of pictureWidth. Consequently, the value of rightOffset would be ‘0’ in such cases when i=num column minus 1+1.
Referring to
In some embodiments, the decoder 700 can receive a coded bitstream 701 into a high level syntax interpreter 702. The bitstream 701 is shown in
The column/slice decoder 705 may be coupled to the high level syntax interpreter 702 and to the parameter set database 703, and may receive the output data 706 from the high level syntax interpreter 702. In some embodiments, the column/slice decoder 705 may decode slice headers, and identify the column to which the first CTB of the slice belongs, for example, by taking into account column boundary information that can be located in parameter sets, stored in the parameter set database 703, and retrieved 704 based on the parameter set reference in the slice header. If the slice/column decoder 705 determines that a slice resides in a single column, as described above, then the column/slice decoder 705 may identify the column in which the slice resides by identifying the column to which the first CTB of the slice belongs.
Based on the column information, the column/slice decoder 705 may select one column decoder for the decoding of the slice. Two column decoders 707, 708 are shown in
Assuming that the slice currently to be decoded is assigned to column decoder 707, data 709 required to decode the column can be forwarded to the column decoder 707. Such data 709 may have similar characteristics as output data 706 from the high level syntax interpreter 702. In some implementations, the column/slice decoder 705 may prune, or mark as irrelevant, those parts of the output data 706 that are not required for the decoding of the column, with the effect that such unneeded parts are not forwarded to column decoder 707 as part of data 709.
In some embodiments, the bitstream 701 may be configured or may otherwise allow for completely or near completely independent decoding of its content in multiple column decoders (e.g., 707 and 708 in
In some embodiments, sharing of information can be handled by a suitably configured sharing unit 710. The sharing unit 701 may be implemented as shared memory, a queuing mechanism that allows message exchange, or other mechanisms for inter-processor or inter-core communication. Where information sharing is enabled by sharing unit 710, it may potentially be necessary to share both meta-information 711 (e.g., entropy decoding state, motion vector predictors) and/or sample data 712 (e.g., values of neighboring pixels of the column used for intra prediction or motion compensation) between the column decoders 707, 708. The sharing unit 710 may also be suitably configured so as to facilitate the “handover” of information from one column decoder 707, 708 to another in such cases where a slice spans across a column boundary between adjacent columns, with the effect that multiple column decoders 707, 708 would be assigned in turn by the slice/column decoder 705 to decode such a slice.
In some embodiments, for example, in implementations of the decoder 700 that utilize a multi-processor or single multicore processor in a personal computer or similar hardware architecture, the implementation of the column decoders 707, 708 as processes running on cores or processors can be greatly facilitated by keeping all state information of the decoder 700, including for example full reference pictures, in memory that is shared between the column decoders 707, 708. In this case, the sharing unit 710 may include a shared memory aided in some cases possibly by interprocessor communication tools such as semaphores or message passing implemented, for example, in an operating system kernel.
The output of the column decoder 708 can be partly reconstructed samples 713 and associated metadata 714 provided to a combiner 715. Likewise the column decoder 707 may produce partly reconstructed samples and associated metadata, similar to partly reconstructed samples 713 and associated metadata 714 produced by column decoder 708. The samples 713 may be described as “partly” reconstructed because for full reconstruction, as specified in the video compression standard, further processing may be required in the picture-based sample generator 717, as described later. The samples 713 and associated metadata output 714 provided by column decoder 708 may be combined in the combiner 715 with the corresponding outputs of the column decoder 707 to generate a column-processed picture 716 and associated metadata 719.
The picture-based sample generator 717 may be configured so as to generate the final reconstructed picture 718 based on the column-processed picture 716 and associated metadata 719, which is received from the combiner 715. The final reconstructed picture 718 maybe used possibly for output by the decoder 700 and/or as a reference picture that is fed back to other components within the decoder 700 (noted further below). The picture-based sample generator 717 can advantageously implement tasks that are not necessarily very computationally intensive, but which require access to information relating to all columns. In-loop filtering, post filtering, and error concealment are examples of those tasks. However, depending on the characteristics of the encoded bitstream 701, it may also be possible that some or all of these tasks can be implemented instead in the column decoders 707, 708 (further details of which may be found in co-pending U.S. patent application Ser. No. 13/336,475). Therefore, in some embodiments, the picture-based sample generator 717 may be omitted from the decoder 700.
Not specifically shown in
Additionally, software based implementations are possible using general purpose processing architectures, an example of which is the data processing systems 800. For example, using a personal computer or similar device (e.g., set-top-box, laptop, mobile device), such an implementation strategy may be possible as described in the following. As shown in
According to various embodiments, the above described method(s) may be implemented by a respective software module. According to other embodiments, the above described method(s) may be implemented by a respective hardware module. According to still other embodiments, the above described method(s) may be implemented by a combination of software and hardware modules.
While the embodiments have, for convenience, been described primarily with reference to an example method, the apparatus discussed above with reference to a data processing system 800 may, according to the described embodiments, be programmed so as to enable the practice of the described method(s). Moreover, an article of manufacture for use with a data processing system 800, such as a pre-recorded storage device or other similar computer readable medium or product including program instructions recorded thereon, may direct the data processing system 800 so as to facilitate the practice of the described method(s). It is understood that such apparatus and articles of manufacture, in addition to the described method(s), all fall within the scope of the described embodiments.
In particular, the sequences of instructions which when executed cause the method described herein to be performed by the data processing system 800 can be contained in a data carrier product according to one embodiment of the invention. This data carrier product can be loaded into and run by the data processing system 800. In addition, the sequences of instructions which when executed cause the method described herein to be performed by the data processing system 800 can be contained in a computer program or software product according to one embodiment of the invention. This computer program or software product can be loaded into and run by the data processing system 800. Moreover, the sequences of instructions which when executed cause the method described herein to be performed by the data processing system 800 can be contained in an integrated circuit product (e.g., a hardware module or modules) which may include a coprocessor or memory according to one embodiment of the invention. This integrated circuit product can be installed in the data processing system 800.
The embodiments of the invention described herein are intended to be exemplary only. Accordingly, various alterations and/or modifications of detail may be made to these embodiments, all of which come within the scope of the invention.
This application claims priority from U.S. Provisional Patent Application Ser. No. 61/427,569, filed Dec. 28, 2010, entitled “PICTURE SEGMENTATION USING GENERALIZED SLICES”, the entire contents of which are incorporated herein by reference.
| Number | Date | Country | |
|---|---|---|---|
| 61427569 | Dec 2010 | US |
| Number | Date | Country | |
|---|---|---|---|
| Parent | 18097019 | Jan 2023 | US |
| Child | 18604341 | US | |
| Parent | 17833115 | Jun 2022 | US |
| Child | 18097019 | US | |
| Parent | 17174728 | Feb 2021 | US |
| Child | 17833115 | US | |
| Parent | 16290764 | Mar 2019 | US |
| Child | 17174728 | US | |
| Parent | 14561480 | Dec 2014 | US |
| Child | 16290764 | US | |
| Parent | 13336675 | Dec 2011 | US |
| Child | 14561480 | US |
