JVET-AG0110 AHG12: 16 States TCQ with State Exchange [M. Balcilar, K. Naser, Y. Chen, F. Galpin, F. Le Léannec (InterDigital)]

This contribution proposes to increase the number of trellis code quantization (TCQ) states in dependent quantization (DQ). The number of states is increased from 8 to 16, where the 8 new states use similar transition rule as the original ones.

The following results {Y, U, V; EncT, DecT, VmPeak} are obtained on top of ECM-11.0:

• RA: { -0.08%, -0.07%, -0.04% ; 100.9%, 99.3%, 101.2%}.
• LDB:{ -0.15%, 0.61%, -0.22% ; 101.0%, 100.3%, 101.6%}.

It was commented by the proponent that the proposed change does not provide good tradeoff for AI, as it introduces around 7-8% encoding run time increase, and 0.1% gain for AI. In the proposal the proposed 16-state TCQ is only enabled for RA and LDB configs. The proponent suggests that this could be due to naïve encoder implementation, and a more sophisticated encoder could potentially achieve better tradeoff for AI.

It was asked by multiple experts if the proposed new states and transitions can be achieved by encoder-only modifications, i.e. if the current 8-state TCQ is capable of achieving these transitions. The proponent claimed that the proposed normative changes are more straightforward.

It was asked for the proponent to show an example of a transition that is not achievable by the current 8-state TCQ. It was commented that the description in the proposal is not sufficient to understand.

It was commented that offline discussion was initiated by non-proponent but not answered by proponent. As a general principle, offline discussion is encouraged as it often facilitates a more efficient meeting discussion.

Multiple experts commented that the performance vs. runtime tradeoff appears interesting.

It was requested for the proponents to provide a description in the EE document (JVET-AG2024) as detailed as possible within the timeline.

It was agreed to investigate this in the EE. Besides providing software and results, proponents are also requested to provide at least one clear example of a transition that is not achievable by the current 8-state TCQ, and get this cross checked.