JVET-AD0163 AHG11: Updating NNPFC and NNPFA SEI messages for the NNPF [M. Santamaria, M. M. Hannuksela, F. Cricri, R. Yang (Nokia)]

This document describes the usage of the most recent NNPFC and NNPFA SEI messages for the NNPF in the context of NNVC 4.0. It is reported that the overall coding gains are -4.92% (Y), -18.36% (Cb), -16.67% (Cr) when all the NN tools are disabled.

It is proposed to update the NNVC software to match with the latest NNPFC and NNPFA SEI messages and the method of applying the base and updated filters as described in this document.

Updates to the implementation of the SEI messages according to latest state (more changes could occur at current meeting).

The model of the filter was not changed – improvements were due to further encoder optimization and reduction of parameter signalling.

Training of models is not included in the encoding time.

It was reported that the base model (which is used to re-train on a per-sequence adaptation level) wsas one of the previous loop filter architecture from Qualcomm which was further generally fine-tuned (before re-training on sequence level). The Qualcomm model was never cross-checked, but the general fine-tuning (as well as sequence adaptive training) was cross-checked by Ericsson in the last meeting. This was agreed to be sufficient as a proof of concept.

The numbers were confirmed by a cross-checker.

Decision (SW/align/encoder): Adopt JVET-AD0163.