JVET-AN0112 Non-EE2: Coding of a coefficient level in TSRC [Y. Yu, L. Xu, J. Gan, H. Yu, D. Wang (OPPO)]

This contribution proposes to always code a significant flag with context to indicate if the coefficient level is zero or not for all coefficient levels within all non-zero subblocks of a TSRC block. In order to guarantee all significant flags coded with context and all context coded bin consumption still follows the current TSRC CABAC bin budget, this contribution proposes to reserve as many context bins as the total number of positions of all non-zero subblocks. The pre-defined CABAC bin budget for a TSRC block is updated accordingly by subtracting the number of reserved context bins and then used as a new CABAC bin budget to control the coding of other context coded syntax. In addition, all significant flags within a subblock are coded together in a separate coding pass before coding any other context coded syntax elements for a coefficient level so that there is no interleaving of context coded bins and bypass coded bins. This contribution also proposes that an adaptive Rice parameter may be used to binarize the remaining of the coefficient levels and Rice parameter is signalled at slice level if needed. For CTC application, adaptive Rice is disabled and Rice parameter 0 is used. On top of ECM-18.0, the simulation results are reported as follows.

AI: Natural content: { 0.00%, 0.00%, 0.03%, 100.1% , 99.8%}

Class F: {-0.12%, -0.15%, -0.27%, 99.1%, 98.1%}

Class TGM: {-0.21%, -0.26%, -0.32%, 99.2%, 100.5%}

RA: Natural content: { -0.01%,,0.01%,,-0.04%, 100.4%, 99.9%}

Class F: {-0.06%, -0.02%, -0.02%, 100.5%, 99.3%}

Class TGM: {-0.16%, -0.16%, -0.08%, 100.3%, 100.6%}

LDB: Natural content: { 0.00%, -0.03%, 0.02%, 100.5%, 100.3%}

Class F: {0.08%, -0.02%, -0.53%, 101.0%, 101.0%}

Class TGM: {-0.17%, -0.07%, -0.20%, 101.0%, 98.2%}

LowQP AI: Natural content: { -0.03%, -0.07%, -0.07%, 100.6%, 100.2%}

Class F: { -0.18%, -0.14%, -0.18%, 99.6%, 100.3%}

Class TGM: { -1.03%, -0.95%, -0.90%, 101.2%, 100.4%}

This proposal is related to JVET-AM0062, which was an EE test at the last meeting.

To solve the performance loss at low QP, adaptive Rice parameter derivation (which is a method supported in VVC v2 but with small modification of the set of allowed parameters) at slice level is used, invoked by encoder-side screen content detection algorithm.

It was asserted by the proponent that if in combination with the encoder modification in JVET-AN0086, this proposal could potentially achieve more gain than reported for low QP. proposal would increase the number of TSRC coefficient coding passes 1 pass in order to code significance flags in a separate pass, which are all context coded. Then, this number is subtracted from the total budget.

Cross checker confirmed the reported results.

Several experts expressed support for this method.

The method proposed in this contribution can be combined and harmonized with JVET-AN0087 and conducted as one EE test for the next round of investigation.

It was commented that JVET-AN0086 was adopted as an encoder software change but turned off by default for CTC.

It was commented that the adaptive Rice parameter derivation in RRC could potentially be used in TSRC to achieve some of the gains reported here. However, that would be a different proposal.

It was agreed to investigate this in an EE, as one EE test together with JVET-AN0087. For low QP results, provide test results should be provided with the JVET-AN0086 encoder change turned on and turned off.