Back to Search Document details
32nd Meeting: Hannover, DE, October 2023 2023-10-13 15:14
EE2 related: CABAC parameters retraining
Abstract
This contribution proposes to update context initialization parameters for all slice types. All CABAC parameters (initial probabilities parameters, window sizes, adaptive weights and rate offsets) are retrained and updated. As previous scripts are not suitable anymore, a set of new scripts taking any account new parameters and constraints is proposed. It is reported that on top of ECM-10.0, the overall coding performance impact for {Y, U, V} is {-0 %, 0.%, 0.% } {-0.%, 0.%, 0.% } {-0.%, 0.%, 0.% } in AI, RA and LDB configurations respectively.
JVET-AF0151 EE2 related: CABAC parameters retraining [F. Galpin, F. Lo Bianco, C. Salmon-Legagneur, K. Naser (InterDigital)]

This contribution proposes to update context initialization parameters for all slice types. All CABAC parameters (initial probabilities parameters, window sizes, adaptive weights and rate offsets) are retrained and updated. As previous scripts are not suitable anymore, a set of new scripts taking any account new parameters and constraints is proposed. It is reported that on top of ECM-10.0, the overall coding performance impact for {Y, U, V} is {-0.16%, -0.03%, -0.10% } {-0.15%, 0.19%, 0.11% } {-0.11%, 0.91%, -0.06% } in AI, RA and LDB configurations respectively.

It was commented that the U component in LDB and LDP show a non-negligible performance loss.

The cross checker used the retrained contexts from the proponent to run the simulation, but didn’t cross check the training due to lack of scripts. It was suggested to investigate this in an EE and include the training scripts in the EE.

It was agreed to investigate this in an EE, and to include the training script in the EE.

Decisions
It was agreed to investigate this in an EE, and to include the training script in the EE.
Citation