JVET-L0051 CE1-related: Partitioning Clean-ups [M. Park, M. W. Park, K. Choi (Samsung)]

This contribution proposes two modifications of the partitioning structure in VVC. One is to remove the adaptive signalling of the maximum binary tree size from the slice header, and the other is to set the maximum binary and ternary tree sizes to the possible maximum sizes by default. The first clean-up reportedly shows –0.08% BD-rate with 107% encoding time on RA. When an encoder optimization technique is applied on top of the first clean-up, the result reportedly shows 0.02% BD-rate with 99% encoding time. The second clean-up reportedly shows –0.57% BD-rate with 140% encoding time on AI, –0.12% BD-rate with 105% encoding time on RA. When an encoder optimization technique is applied on top of the second clean-up, the results reportedly show –0.05% for Y, –0.87% for U and –0.86% for V BD-rates with 102% encoding time on AI, and –0.01% for Y, –0.39% for U and –0.35% for V BD-rates with 101% encoding time on RA. Results of two clean-ups with the encoder optimization techniques reportedly show –0.05% for Y, –0.87% for U and –0.86% for V BD-rates with 102% encoding time on AI, and 0.02% for Y, –0.37% for U and –0.33% for V BD-rates with 99% encoding time on RA.

Removal is not a good option, as encoders can use this beneficially (e.g. if they don’t use the maximum depth in their checks).

It is furthermore requested to use identical maximum depth for intra and inter slices, and for intra and inter. By increasing the maximum depth for intra slices, the above mentioned gain of approx. 0.6% (at 140% encoding time) is achieved (“second cleanup”), but loss for RA.

The encoder optimization technique reduces the runtime, but the gain is almost gone.

The current signalling of VVC allows signalling the max size of binary split only for inter, whereas it is fixed to 32 for luma and 64 for chroma. The proposal shows that compression benefit is possible by increasing the value to 64 for luma as well (which is not the same for inter, where it is 128). Allowing a maximum BT size 64 for luma in intra slices seems desirable, whereas CTC should stay with 32 (to avoid increase of runtime). Another proposal (L0218) requests such a signalling.

Currently, the max size of ternary split is fixed to 32 for intra, 64 for inter. The proposal requests changing it from 32 to 64 for intra; as it would be desirable to retain 32 for CTC, it would be necessary to define a signalling mechanism for max ternary tree split size as well. This signalling is not requested in this proposal, and no results are available for this.

No action was taken on this specific proposal.