JVET-L0361 CE1-related: Context modeling of CU split modes [Y. Zhao, H. Yang, J. Chen (Huawei)] [late]
This contribution proposes a modified context modeling for CU partitioning flag signalling. The proposed method achieves −0.06%, −0.12%, and −0.20% luma BD-rates for the AI, RA, and LB settings of CTC, respectively.
Two versions:
- increasing number of context models from 17 to 19 gives 0.0%, 0.10%, 0.16% luma BR red. for AI/RA/LD
- increasing number further to 22 gives 0.06%/0.12%/0.20% luma BR red. for AI/RA/LD
Some experts expressed this is not adding significant complexity (no new context coded bins, only more complex models which need some additional storage).
Concern was expressed by proponents of JVET-K0362 (part of which is included here), and by cross-checker. It was for example mentioned that the proposal requires additional checks depending on block size threshold to determine which context model would be applied. These concerns were later resolved, abd therefore consensus was reached to adopt the proposal (follow-up in Track A Tue afternoon)
Decision: Adopt JVET-L0361 (version with 22 context models).