JVET-AM0185 AhG11: Decomposed Content-Adaptive VLOP [Z. Xu, J. Konieczny, A. Filippov, C. Hollmann, V. Rufitskiy, T. Dong, H. Qin (TCL)]

This contribution studies the dimension-wise decomposed representation of multiplier in case of content-adaptive VLOP (CAVLOP). The BD-rate gain under RA configuration are reported as below.

Compared with NNVC-13 with CAVLOP ON, the averaged gain is reported to be:

{-0.09 %, -0.40%, -0.32%, EncT: XX.X%, DecT: XX.X%}

Previously, the decomposed approach was considered for LOP but not adopted, as the training time increase (which is critical for content adaptation) did not justify the additional compression benefit. The proponents believe that it is almost the same training time as for existing VLOP, as in case of VLOP less parameters are trained.

It was agreed to investigate this in an EE. Comparison of using NNVC14 VLOP for content-adaptive filtering, and compare it against the dimension-wise decomposed approach (also with attention mechanism in case that EE1-2.1 becomes VLOP4).