JVET-AF0206 AHG11: Complexity reduction of NN-based loop-filters [J. Li, Y. Li, C. Lin, K. Zhang, L. Zhang (Bytedance)]

Objection was raised against presentation of this document on Wednesday Oct. 18, as the first version with reasonable description and results only was made available by Oct. 17.

This contribution proposes a unified filter architecture, with reported complexity of 5.0 and 7.3 kmac/pixel in two tests, which is even lower than 17.0 kmax/pixel of the lower operation point (LOP) in NNVC. The training and inference are kept same as the existing NN-based filters in NNVC.

Compared with NNVC-6.0 VTM anchor the BD-rate changes are shown as below reportedly:

Test #1: AI: -3.14%, -6.34%, -5.29%, RA: -3.09%, -6.22%, -4.04% (On top of NNVC-6.0 VTM anchor, 7.3 kmac/pixel)

Test #2: AI: -2.47%, -4.63%, -4.21%, RA: -1.96%, -3.98%, -2.93% (On top of NNVC-6.0 VTM anchor, 5.0 kmac/pixel)

This was presented Thu. Oct. 18, at 1155.

It was commented that it is questionable whether even lower complexity operation point would be needed at this stage of the exploration, and whether otherwise this contribution provides more insight.

It was commented that RA has a significant drop compared to LOP, and that according to another experts’ experiments, better performance could be achieved.

Further study was recommended.