JVET-AE0147 AHG12: Updating context model parameters for ECM [R.-L. Liao, Y. Ye, J. Chen, X. Li (Alibaba)]

This contribution proposes to update context initialization parameters for I-slice. Specifically, the initial probabilities, window sizes and adaptive weights for I-slice context models are retrained and updated. It is reported that on top of ECM-9.0, the overall coding performance impact for {Y, U, V} is {-0.10%, 0.07%, 0.15%} for AI, {-0.01%, -0.01%, 0.03%} for RA and {0.05%, -0.22%, 0.32%} for LB.

The training script was used that was once used during in VVC development, and adapted to ECM accordingly. It was reported that currently, the script is not properly working for PB slices, likely due to temporal inheritance. Contexts are also more complex in the ECM, so the modifications necessary may not be trivial.

It was questioned whether the gain in intra is justified, while loss occurs in LB. It is also observed that the gain is quite sequence dependent, some sequences with high.

It was reported that in case of VTM, also other content such as HDR was used.

Several experts expressed that it would be beneficial to update initialization regularly. However, the proposed approach does not seem to do it in an optimum way, and in particular not for PB slices yet.

It was agreed to study this in the EE, provided that the modified script could be shared and if possible further improved.

Further study was recommended for the application to PB slices.