Tool Experiment 2 on IBDI and memory compression - DOC

Document Sample
Tool Experiment 2 on IBDI and memory compression - DOC Powered By Docstoc
					Joint Collaborative Team on Video Coding (JCT-VC)
                                                                           Document: JCTVC-B302r2
of ITU-T SG16 WP3 and ISO/IEC JTC1/SC29/WG11
2nd Meeting: Geneva, CH, 21-28 July, 2010

Title:             Tool Experiment 2 on IBDI and memory compression
Status:            Output Document Approved by JCT-VC
Purpose:           TE description
Author(s) or       Takeshi Chujoh
Contact(s):                                                       Tel:     +81-44-549-2288
                   1, Komukai-Toshiba-cho, Saiwai-ku,             Email:
                   Kawasaki, 212-8582, Japan
Source:            TE Coordinator

1. Introduction
The goal of this Tool Experiment (TE) is to further investigate IBDI and memory compression
technology. This TE is an ongoing work and results of previous TE2 in JCTVC-A302r1 are summarized
in JCTVC-B046r2. In this stage, software platform and test condition are updated and by using
measurement module, each proposal will be measured more specific results and analyzed from many

2. Participants
Coordinator: Takeshi Chujoh (TOSHIBA Corporation)
         No.    Affiliation            Contact                       Email                   Remark
          1        NEC              Hirofumi Aoki                 Proponent
          2      Panasonic       ChongSoon Lim Proponent
          3       SHARP             Andrew Segall             Proponent
                   Texas        Madhukar Budagavi     
          4                                                                                  Proponent
                Instruments       Minhua Zhou           
          5     TOSHIBA             Takeshi Chujoh      Proponent
                 Yonsei             Yoonsik Choe                  Cross-
                University          Soongi Hong                    checker
          7        Intel             Yi-jen Chiu    
          8        JVC               Ichiro Ando  
                                 Seungwook Park   
          9         LG
                                   Sehoon Yea       
         10     Qualcomm             Rajan Joshi   
                                  Ali Tabatabi    
         11        Sony
                                 Teruhiko Suzuki    
         12    Zenverge, Inc.       Dzung Hoang  

                                                 Page: 1                              Date Saved: 2011-01-13
                                    Lu Yu          
      13                           Jie Chen    
                                 Xiaolin Shen    

3. Functionality addressed
    • Improvement of coding efficiency by increasing internal process of video codec while
       minimizing memory access bandwidth
    • Reduction of reference frame memory access bandwidth and reference frame memory size while
       minimizing visual quality degradation.

4. Description of tool
Details are provided in JCTVC-B044, JCTVC-B057, JCTVC-B089, JCTVC-B103 and JCTVC-B114.

4.1. JCTVC-B089 (Texas Instruments)
An in-loop memory access bandwidth reduction technique is proposed. Proposed tool compresses the
reference pictures before they are written to the memory and decompresses before they are being read.
Fixed compression ratio is targeted for a block of pixels to enable random access. Proposed technique
provides 12 bit/pixel to 8 bit/pixel compression when Internal Bit-Depth Increase (IBDI) tool is turned
on. 8 bit/pixel to 4 bit/pixel compression is achieved when IBDI is disabled. 2-D integer S-transform on
8x8 blocks, DC prediction, quantization and variable length entropy coding is employed.

4.2. JCTVC-B044 (TOSHIBA Corporation)
A bit depth compression method on IBDI is proposed [4]. IBDI is a coding tool that increases the bit
depth of input picture at encoder side and decreases the bit depth of output picture to input bit depth. By
increasing bit depth, the coding efficiency is improved; however, DPB is increased memory capacity and
memory bandwidth compared to storing input depth. This adaptive scaling algorithm is one of the
solutions for this problem. In this TE, we test to compress depth from 12-bit to 8-bit on IBDI and also
integrate this method and in-loop/de-blocking filtering process.

4.3. JCTVC-B057 (NEC Corporation)
In the algorithm proposed in JCTVC-B057, reference frames are compressed before being stored in
memory and they are decompressed after being read from the memory. The proposed method employs a
1-D DPCM for the compression, considering the balance between memory accessibility and compression
efficiency. The 1-D data structure is effective in suppressing overhead caused by unaligned memory
access. DPCM provides decent compression efficiency with sufficiently low complexity and guarantees
fixed compression ratio without any rate control. Figure 1 shows an example for memory compression
with the proposed 1-D DPCM. Here, a 1-D DPCM block consists of eight consecutive pixels. The left-
most pixel is first coded with PCM and serves as a base point. The other seven pixels are then coded with
DPCM from left to right, using a five-bit nonlinear quantizer. Actually the block size, i.e., base-pixel
interval and bit allocation for prediction errors can be variable in addition to a nonlinear quantization
matrix. These parameters are signaled to decoders as side information for each sequence parameter set.
Details including algorithm description and bitstream syntax are described in [5].

                                            Page: 2                               Date Saved: 2011-01-13
                      8 8 8 8 8 8 8 8 8 8 8 8 8 8 8 8

                                              Uncompressed image


                      8 5 5 5 5 5 5 5 8 5 5 5 5 5 5 5

                                                Compressed data

                                            DPCM block
                                      B     B-bit base pixel with original value
                                      M     M-bit prediction error pixel with quantized value
                                            Prediction direction

                      Figure 1 1-D DPCM prediction structure and bit allocation.

4.4. JCTVC-B103 (Panasonic)
In this contribution, a simple image coder based reference frame compression scheme is presented.

 A Block Of Image
                          Transform                 Scanning                 Entropy Coding              Data Storage

                            Figure 2        Reference Frame Compression Scheme

 A Block Of Image
                       Inverse Transform         Inverse Scanning           Entropy Decoding             Data Storage

                           Figure 3        Reference Frame Decompression Scheme

The compression scheme described in Figure 2 is similar to a traditional image coding scheme except that
quantization process is purposely omitted. A quantization process in an image compression scheme is
used typically to control the compression rate. The reason for omitting the quantization process is
because to regulate the final size of the compressed data to the desired data size will require some form of
rate control. It is usually difficult to estimate the quantization parameter for the quantization process so
that the output size of the entropy coding process is within the targeted size in just one single encoding
pass. Therefore to replace the function of the quantization process, bit plane coding is used for entropy
coding. Using bit plane coding, the desired compression rate can be easily achieved by fitting the coded
bits to the available data storage size in order of significance. The information that is not able to be
fitted within the available data storage size is discarded and can be considered as “quantized”.

                                               Page: 3                                         Date Saved: 2011-01-13
The transform used in JCTVC-B103 is a floating point DCT. A Hadamard transform will be used in the
Tool experiment. The bit plane coding used in this contribution is described below.

                             Order Of Bits To Be Stored In Data Storage Until Data
                                               Storage Is Filled

        Number Of Bit
                             Bit Plane 1                Bit Plane 2                         Bit Plane N
         Planes, N                                                         *       *

            u( 4 )

                End Of                      End Of                        End Of               End Of
                            Run, Signed                 Run, Signed                Run, Signed            Other Significant Bits
                BitPlane                    BitPlane                      BitPlane
                             Bit Pair 1
                                                         Bit Pair 2   *   *Flag Bit Pair Kn BitPlane
                                                                                                          Of Higher Bit Planes

                     u(1)                      u(1)                         u(1)                u(1)
                                                                                                                 u( m )
                                                                                                                    n 1
                                                                                                             m  Kj
                                                                                                                    j 1


                            n( v )         u( 1 )

                                 Figure 4             Bit Stream Structure Of One Data Block.

Figure 4 shows the coding order of bits to be stored in the data storage. The bits are coded and decoded
in the same order. The bits that cannot be stored into the data storage are discarded.
As shown in figure, a value representing the number of bit planes is coded in 4 bits. It indicates the
minimum number of bit planes required to represent the largest absolute values of the coefficients.
Next, each bit plane is coded starting from the bit plane containing the most significant bit of the largest
absolute value. The coding of each bit plane is as follows:
Firstly, a flag indicating if it is the end of the bit plane is coded using one bit. This is followed by the
coding of run and signed bit pair. For each run and signed bit pair, the run parameter indicates the
position of a Most Significant Bit and it is coded using unarycode and the signed bit parameter is coded
using one bit. The coding of each run and signed bit pair is preceded with coding of an end of plane
When the end of plane flag is coded with the value one, it indicates there is no more Most Significant Bit
existing in the bit plane starting from the last coded position. After the coding of end of plane flag with
a value of one, the significant bits of other positions, whose most significant bits have previously
occurred in higher bit planes, are coded using one bit per significant bit.
Thus the entire entropy coding only uses variable length codes for the run parameter. For the most
significant bit plane, the first bit to indicate the end of plane is not required to be coded.

                                                           Page: 4                                        Date Saved: 2011-01-13
4.5. JCTVC-B114 (SHARP)
We propose a system that enables the low resolution decoding of a bit-stream without drift. The system
consists of a buffer compression algorithm that reduces memory bandwidth for all devices, and a low
resolution decoding mode that enables optional, lower power operation. We assert that this lowest
power mode is beneficial for battery powered devices and additionally benefits devices with screen
resolutions lower than the content resolution. To achieve our goal, we propose to store sub-sampled and
compressed versions of reconstructed frames in the decoded picture buffer. We then store prediction
and residual data to reconstruct the missing pixels. The result is buffer compression.
At the high level, our encoding technique has the following steps: We begin with a reconstructed and
full resolution image frame. To be clear, this is the output of the HEVC decoder. Then, we store a
low-resolution version of the full resolution frame. This low-resolution version consists of half the
image pixels and corresponds to a quincunx sampling of the pixel data as shown in Figure 5. Note that
we do not filter the data. Next, this low-resolution version is coded using the well known adaptive
moment block truncation coding (AMBTC) technique and stored in the frame buffer. Subsequently, we
predict the missing pixels from the stored values using bi-linear interpolation. The residual differences
are then quantized and stored using the same technique.

          Figure 5 - Sampling structure of the frame buffer compression algorithm
A diagram of the approach appears in Figure 6. As can be seen from the Figure, the decoding process
reverses that of the encoder. Namely, we fetch the down-sampled data and residual information. Then,
we predict the missing pixels from the down-sampled data and add the residual.

                 Removal                      Pixel
                                          Reconstruction        U
                                                            L   C   R
                                                    C = a1*L + a2*U + a3*R +       Checkboard
                                                              a4*B                Reconstruction
 Figure 6 - Block diagram of the integration of the frame buffer compression algorithm in
                                      the video codec.

5. Expected gains
    • Improvement of coding efficiency by increasing internal process of video codec while
       minimizing memory access bandwidth
    • Reduction of reference frame memory access bandwidth and reference frame memory size while
       minimizing visual quality degradation.

                                           Page: 5                               Date Saved: 2011-01-13
6. Tool experiments conditions

6.1. Software
TMuC version 0.7 (version Aug. 11 2010)

6.2. Coding conditions
Depth compression ratio 1: 12-bit to N-bit (Anchor including IBDI)
  High coding efficiency, random access configuration
  High coding efficiency, low delay configuration

Depth compression ratio: 8-bit to N-bit
  Low complexity, random access configuration
  Low complexity, low delay configuration

Common test condition is specified in JCTVC-B300 [8].

6.3. Evaluation criteria
   1.   Measure impact on bitrate/PSNR using provided data. Use 4-point BD-PSNR and BD-Rate.
   2.   Memory compression ratio with stored format and controlling method
   3.   Memory access measures (Measurement module will be provided by some proponents)
   4.   Complexity (encoding and decoding times)
   5.   Subjective quality (informal comments)

6.4. Time lines
Aug.13 2010     : Deadline for sending email to coordinator expressing interest in participating in tool
Aug.13 2010     : Deadline of formal description of all proposals.
Aug.13 2010     : Distribution of document of tool experiment.
Sept.3 2010     : Provide measurement module
Sept.15 2010    : Start to cross-check of experimental results and check subjective quality
Sept. 24 2010   : Complete cross-checking of experimental results and checking subjective quality
Oct.1 2010      : Upload contributions

6.5. Reference
[1] T. Chujoh, “Tool Experiment 2 on IBDI and memory compression,” JCTVC-A302r1, 1st. Meeting,
Dresden, DE, 15-25 April, 2010.
[2] T. Chujoh, “Summary of Tool Experiment 2 on IBDI and memory compression,” JCTVC-B046r2,
2nd. Meeting, Geneva, CH, 21-28 July, 2010.
[3] M. U. Demircin, M. Budagavi, M. Zhou and S. Dikbas, “TE2: Compressed reference frame buffers
(CRFB),” JCTVC-B089, 2nd. Meeting, Geneva, CH, 21-28 July, 2010.

                                            Page: 6                                Date Saved: 2011-01-13
[4] T. Chujoh, T. Shiodera and T. Yamakage, “TE2: Adaptive scaling for bit depth compression on
IBDI,” JCTVC-B044, 2nd. Meeting, Geneva, CH, 21-28 July, 2010.
[5] H. Aoki, K. Chono, K. Senzaki, J. Tajime and Y. Senda, “Performance report of DPCM-based
memory compression on TE2,” JCT-VC document JCTVC-B057, Joint Collaborative Team on Video
Coding (JCT-VC) of ITU-T SG 16 WP3 and ISO/IEC JTC1/SC29/WG11 2nd meeting, Geneva, 2010.
[6] C.S. Lim, H.W. Sun and V. Wahadaniah, “Reference Frame Compression Using Image Coder,” JCT-
VC document JCTVC-B103, Joint Collaborative Team on Video Coding (JCT-VC) of ITU-T SG 16 WP3
and ISO/IEC JTC1/SC29/WG11 2nd meeting, Geneva, 2010.
[7] Z. Ma and A. Segall, “System for Graceful Power Degradation,” JCT-VC document JCTVC-B103,
Joint Collaborative Team on Video Coding (JCT-VC) of ITU-T SG 16 WP3 and ISO/IEC
JTC1/SC29/WG11 2nd meeting, Geneva, 2010.
[8] F. Bossen, “Common test conditions and software reference configurations,” JCT-VC document
JCTVC-B300, Joint Collaborative Team on Video Coding (JCT-VC) of ITU-T SG 16 WP3 and ISO/IEC
JTC1/SC29/WG11 2nd meeting, Geneva, 2010.

                                          Page: 7                             Date Saved: 2011-01-13

Jun Wang Jun Wang Dr
About Some of Those documents come from internet for research purpose,if you have the copyrights of one of them,tell me by mail you!