JVET-AE0176 EE2-related: Enhancements on CCRM [Z. Deng, K. Zhang, L. Zhang (Bytedance)]

This proposal presents the results of a multi-model CCRM mode and a CCRM merge mode. In EE2-3.1a, a single-model CCRM is investigated. In this contribution, two methods are presented to further improve the coding efficiency of CCRM. In method #1, a multi-model CCRM mode is proposed. Samples are separated into two groups and each group can apply its own CCRM model. There is no flag signalling involved in the proposed method, and the final CCRM mode is implicitly determined by comparing decoder derived costs of the two CCRM modes with single or multiple models. In method #2, a CCRM merge mode is proposed, wherein the CCRM model is inherited from a candidate in the CCRM merge list. Simulation results based on ECM-9.0 are reported as below:

Method #1 :

RA { -0.20%, -1.66%, -1.89%, 101.8%, 100.0% }; LDB { -0.01%, -3.86%, -3.46%, 101.9%, 99.7% }

Method #2 :

RA {}; LDB {}

Simulation results versus EE2-3.1a are listed as below :

Method #1 :

RA { -0.03%, -0.07%, -0.07%, 100.4%, 100.1% }; LDB { -0.12%, 0.39%, 0.42%, 100.7%, 99.8% }

Method #2 :

RA {}; LDB {}

Relative to EE test 3.1a, small benefit in RA; in LB, some gain in luma, but loss in chroma.

According to proponents of 3.1, it would be a straightforward add-on, and might be interesting to investigate on top of the adopted method 3.1b.

It was agreed to investigate this in the upcoming EE.