Joint Video Experts Team (jvet) of itu-t sg 6 wp and iso/iec jtc 1/sc 29/wg 11

7.7CE7 related – Quantization and coefficient coding (7)

Contributions in this category were discussed Saturday 14 July 1045–XXXX (chaired by GJS).

JVET-K0070 Non-CE7: Entropy Coding for Dependent Quantization [M. . Coban, J. . Dong, M. . Karczewicz (Qualcomm)]

In this document an alternate dependent quantization entropy coding scheme that allows grouping of bypass coded bins at coefficient group level is described. The state machine driving the quantizer selection and coefficient coding algorithm is driven by the significance of a coefficient rather than the coefficient level parity allowing grouping of remaining level bypass coded bins for efficient parsing. The simulations results reportedly show 4.41% AI, 3.01% RA, and 2.44% LB BD-Rate gains versus the VTM-1.0 anchor.

To be studied in a CE.

JVET-K0424 Cross-check of JVET-K0070: Non-CE7: Entropy Coding for Dependent Quantization [H. . Schwarz (Fraunhofer HHI)] [late]
JVET-K0072 Non-CE7: Alternative entropy coding for dependent quantization [H. . Schwarz, T. . Nguyen, D. . Marpe, T. . Wiegand (Fraunhofer HHI)]

This contribution describes an alternative approach for entropy coding of transform coefficient levels for dependent quantization. It is asserted that the proposed entropy coding can be used with dependent scalar quantization and conventional scalar quantization, only a single 16-bit parameter (which represents the state transition table for dependent quantization) has to be set depending on the quantization approach used. The proposal uses exactly the same concept of dependent quantization as JVET-K0071 (tested in CE7-2.1). In comparison to the transform coefficient coding in JVET-K0071, the transform coefficient levels of a subblock are coded using multiple passes over the scan positions, where all bypass-coded bins of a subblock are transmitted after the regular-coded bins. The proponents assert that the bin-to-bin dependencies are reduced relative to the approach in JVET-K0071 and that the alternative approach is more suitable for high-throughput hardware implementations.

The following average results bit rate reductions are reported relative to VTM-1:

Coefficient coding only:

• 
  AI: −1.6%, −1.5%, −1.7% (Y, Cb, Cr) at 105% encoder and 102% decoder run time;
  
• 
  RA: −1.0%, −1.0%, −1.0% (Y, Cb, Cr) at 104% encoder and 100% decoder run time;
  
• 
  LB: −0.7%, −1.6%, −0.4% (Y, Cb, Cr) at 104% encoder and 101% decoder run time;
  
• 
  LP: −0.8%, −1.7%, −0.3% (Y, Cb, Cr) at 104% encoder and 99% decoder run time.
  

Combination of coefficient coding and dependent quantization:

• 
  AI: −4.8%, −1.6%, −1.6% (Y, Cb, Cr) at 119% encoder and 100% decoder run time;
  
• 
  RA: −3.2%, −1.5%, −1.5% (Y, Cb, Cr) at 113% encoder and 99% decoder run time;
  
• 
  LB: −2.6%, −2.8%, −0.9% (Y, Cb, Cr) at 116% encoder and 100% decoder run time;
  
• 
  LP: −2.4%, −2.7%, −1.0% (Y, Cb, Cr) at 118% encoder and 98% decoder run time.
  

Encoder runtimes are increased about 15-20%.

It was discussed whether sign data hiding can be combined with the TCQ dependent quantization schemes. Further study of this is encouraged.

As a fallback mode, the state aspects can be disabled with ordinary quantization.

It was suggested to put the fallback mode switch at the picture level.

Decision: Adopt (to WD/VTM) with fallback switch for the state aspects at the picture level.

JVET-K0428 Cross-check of JVET-K0072: Non-CE7: Alternative entropy coding for dependent quantization [M. . Coban (Qualcomm)] [late]
JVET-K0319 CE7-related: TCQ with high throughput coefficient coding [J. . Dong, M. . Coban, M. . Karczewicz (Qualcomm)]

This is another proposed variation of TCQ (a.k.a. dependent quantization). This was trying to combine the JEM coefficient coding with TCQ. Detailed presentation was not requested, although the proponent may wish to include it in the planned CE.

JVET-K0513 Cross-check of JVET-K0319: CE7-related: TCQ with high throughput coefficient coding [L. . Zhang (Bytedance)] [late]
JVET-K0281 Inter-Component Context Modelling for Coded Block Flag [C. . Rudat, P. . Haase, H. . Schwarz, D. . Marpe, T. . Wiegand (HHI)]

This is reporting the gain for a change in the context modelling for the coded block flag, making the second chroma CBF dependent on the first one, with a reported gain of about 0.2% for AI and 0.1% for RA. Detailed presentation was not requested.

JVET-K0310 AHG13: Sign Data Hiding [P. . Yin, S. . McCarthy, F. . Pu, T. . Lu, W. . Husak, T. . Chen (Dolby), T. . Tsukuba (Sony)]

It was said that sign data hiding can provide 0.9% on AI and 0.7% on RA for the VTM with 3-–4% encoder increase (and no decoder complexity increase). See also the AHG13 report, which indicated that the previous non-adoption inclusion of SDH in VVC had a larger impact on coding efficiency than other removed features relative to HEVC (other having 0.4% impact or less, and in RA it was the only one with a significant impact), and had a good tradeoff of coding efficiency and complexity.

Decision: Adopt SDH into WD/VTM (can only be used when TCQ dependent quantization is disabled).

7.8CE8 related – Current picture referencing (1)

Contributions in this category were discussed Saturday 14 July 1145–XXXX (chaired by GJS).

JVET-K0050 CE8 related: Intra Region-based Template Matching for luma and chroma [G. . Venugopal, K. . Müller, H. . Schwarz, D. . Marpe, T. . Wiegand (HHI)]

This document proposes the region-based template matching method for both luma and chroma for intra prediction. The proposed method is an extension of the algorithm explained in the JVET document JVET-K0048 to the chroma components. The experimental results of the RTM method for luma and chroma for the all intra (AI) configuration against VTM-1.0 shows an overall BD-rate gain of 2.30% with 120% and 215% decoder and encoder run-time respectively. For the random access (RA) case, the overall coding gain is 0.76% with 103% decoder run-time and 117% encoder run-time, and for the low delay (LB) case it is 0.18% coding gain with 101% decoder run-time and 109% encoder run-time. While using BMS-1.0 as the anchor, the overall BD-rate gain is 1.32% for the all intra (AI) configuration with 104% decoder complexity and 113% encoder complexity. For the random access (RA) configuration, the overall coding gain is 0.43% with 101% decoder run-time and 105% encoder run-time, and for the low delay (LB) case it is 0.16% overall gain with 100% decoder run-time and 102% encoder run-time.

This extends a prior proposal to support chroma, using displacement vectors derived from luma.

This provides more gain on the CTC (2.3% for AI, 0.8% for RA) than other CPR proposals (1.3% for AI, 0.5% for RA), although it provides less gain for SCC than other CPR methods. It uses a decoder template matching search.

Further study is encouraged. Complexity issues need to be considered.

JVET-K0401 Crosscheck of JVET-K0050: CE8 related: Intra Region-based Template Matching for luma and chroma [X. . Xu, X. . Li (Tencent)] [late]