Back to Search Document details
14th Meeting: Geneva, March 2019 2019-03-19 04:09
CE7-related: TB-level constraints on context-coded bins for coefficient coding
Authors: T.-D. Chuang MediaTek S.-T. Hsiang MediaTek Z.-Y. Lin MediaTek C.-Y. Chen MediaTek Y.-W. Huang MediaTek Shawmin Lei MediaTek
Abstract
In the VTM4.0 coefficient coding, a coefficient subblock-level maximum number of context-coded bins is imposed. The averaged maximum number of context-coded bins per 16 samples is increased by 28% compared with that in HEVC. In this contribution, it is proposed to use a transform block (TB) level constraint for maximum number of context-coded bins instead of coefficient subblock-level constraint. For a luma TB, the maximum allowed number of context-coded bin is equal to tb_width*tb_height*1.75. For a chroma TB, the maximum allowed number of context-coded bin is equal to tb_width*tb_height*1.25. The averaged maximum number of context-coded bins per 16 samples is only increased by 1.3% compared with that in HEVC. Simulation results reportedly show that the average luma BD-rates are 0.03%, 0.01%, and 0.07% in AI, RA, and LB cases, respectively.
Meeting Notes
JVET-N0091 CE7-related: TB-level constraints on context-coded bins for coefficient coding [T.-D. Chuang, S.-T. Hsiang, Z.-Y. Lin, C.-Y. Chen, Y.-W. Huang, S.-M. Lei (MediaTek)]

In the VTM4.0 coefficient coding, a coefficient subblock-level maximum number of context-coded bins is imposed. The averaged maximum number of context-coded bins per 16 samples is increased by 28% compared with that in HEVC. In this contribution, it is proposed to use a transform block (TB) level constraint for maximum number of context-coded bins instead of coefficient subblock-level constraint. For a luma TB, the maximum allowed number of context-coded bin is equal to tb_width*tb_height*1.75. For a chroma TB, the maximum allowed number of context-coded bin is equal to tb_width*tb_height*1.25. The averaged maximum number of context-coded bins per 16 samples is only increased by 1.3% compared with that in HEVC. Simulation results reportedly show that the average luma BD-rates are 0.03%, 0.01%, and 0.07% in AI, RA, and LB cases, respectively.

Was also tested at low QPs, does also not introduce loss.

Generally agreed that an approach in this direction is reasonable, as it can be expected that the number of context coded bins is unequally distributed in case of large TUs.

It is suggested to study this in CE. Also investigate the effect on number of context coded bins using the software of JVET-N0049. The CE should also investigate which percentage of overall context coded bins is consumed by the residual coding (in dependency of QP). Low QP range shall be investigated as well. For this purpose, it may not be necessary to study LB performance. The CE should also impose the same constraint for the TS case (which has a constraint at block level, which needs to be modified to the same number)

References:
JVET-M1010
Cited:
JVET-N0027 X JVET-N0106 JVET-N0638 X JVET-N1027
ARCHIVE v3 :
JVET-N0091-v1.docx JVET-N0091-v2-clean.docx JVET-N0091-v2.docx JVET-N0091_DraftText.docx JVET-N0091_Presentation.pdf JVET-N0091_Results_Table2.xlsm JVET-N0091_Results_Table3.xlsm
Decisions
It is suggested to study this in CE. Also investigate the effect on number of context coded bins using the software of JVET-N0049. The CE should also investigate which percentage of overall context coded bins is consumed by the residual coding (in dependency of QP). Low QP range shall be investigated as well. For this purpose, it may not be necessary to study LB performance. The CE should also impose the same constraint for the TS case (which has a constraint at block level, which needs to be modified to the same number)
Citation