Back to Search Document details
22nd Meeting: by teleconference, April 2021 2021-04-21 17:22
AHG8: Content Adaptive Transform Precision for High Bit Depth Coding
Authors: Karam Naser InterDigital Franck Galpin InterDigital Fabrice Le Léannec InterDigital Philippe de Lagrange InterDigital
Abstract
The intermediate transform coefficient precision is fixed to 16 bit signed representation in VVC. Therefore, a pre-computed down-shift is performed after inverse quantization and inverse transform in order to avoid exceeding the 16 bit buffer. It is argued that this down-shift is too conservative as the prediction mechanism results in low magnitude residuals. For High Bit-Depth coding, the down-shift is increased to ensure 16 bit representation, which further decreases the quality of the inverse transformed coefficients. This contribution proposes to adaptively assign the down-shift according the contents.
Meeting Notes
JVET-V0133 AHG8: Content Adaptive Transform Precision for High Bit Depth Coding [K. Naser, F. Galpin, F. Le Léannec, P. De Lagrange (InterDigital)]

The intermediate transform coefficient precision is fixed to 16 bit signed representation in VVC. Therefore, a pre-computed down-shift is performed after inverse quantization and inverse transform in order to avoid exceeding the 16 bit buffer. It is argued that this down-shift is too conservative as the prediction mechanism results in low magnitude residuals. For High Bit-Depth coding, the down-shift is increased to ensure 16 bit representation, which further decreases the quality of the inverse transformed coefficients. This contribution proposes to adaptively assign the down-shift according the contents.

The following experimental results of bdrate gain are reported, for CATP0, Low QP:

  • 0.00%, -0.12% and -0.06% in bdrateY in PQ, HLG and SVT in AI
  • -0.22%, -0.18% and -0.08% in bdrateY in PQ, HLG and SVT in LDB
  • X.XX%, X.XX% and X.XX% in bdrateY in PQ, HLG and SVT in RA

Normal QP

  • -0.04% and 0.00% in bdrateY in H1 and H2 in AI
  • - 0.17% and X.XX% in bdrateY in H1 and H2 in RA

Additional results Low QP, SVT 16 with TrDynamicRange = bitDepth

  • -0.41% for SVT16 in AI

CATP1: Low QP

  • -0.50%, -0.20% and -0.11% in bdrateY in PQ, HLG and SVT in AI
  • -0.39%, -0.28% and -0.13% in bdrateY in PQ, HLG and SVT in LDB
  • X.XX%, X.XX% and X.XX% in bdrateY in PQ, HLG and SVT in RA

Normal QP

  • -0.05% and 0.00% in bdrateY in H1 and H2 in AI
  • - 0.15% and X.XX% in bdrateY in H1 and H2 in RA

Additional results Low QP, SVT 16 with TrDynamicRange = bitDepth

  • -1.50% for SVT16 in AI

Presented in session 13 1415-1435 UTC (chaired by JRO)

The approach is to analyse the maximum value of transform coefficient in current block, and determine the amount of shift. This keeps the buffer size lower, and also less precision in multiplier.

Questions:

  • Would the gain still appear with precision extension flag enabled? Likely not.
  • Was it tested with the CE base anchor? No.

Implementing precision extension flag increases complexity and requires some re-design. The same is true for this proposal, but more investigation would be necessary to assess if it is more implementation friendly.

It was agreed to study this in a CE.

References:
JVET-U2018
Cited:
JVET-V0008 JVET-V0166 X JVET-V2022 JVET-W0022 X JVET-W0050 JVET-W0092
PATENTS:
113796076 0.34 WO/2020/198352 0.33 20220132123 0.32 3949404 0.31 119754754 0.30 1389875 0.28 117668766 0.27
ARCHIVE v2 :
CATP_0_DynamicShift_0_vs_VTM_low-QP.xlsm CATP_0_vs_VTM_low-QP.xlsm CATP_0_vs_VTM_normal-QP.xlsm CATP_2_DynamicShift_0_vs_VTM_low-QP.xlsm CATP_2_vs_VTM_low-QP.xlsm CATP_2_vs_VTM_normal-QP.xlsm JVET-V0133-r1.docx
Decisions
It was agreed to study this in a CE.
Citation