JVET-AM0186 AhG11: Content-Adaptive Neural Network-based Super Resolution [Z. Xu, J. Konieczny, A. Filippov (TCL)]

This contribution studies content adaptation per random-access segment in case of Neural Network-based Super Resolution. The BD-rate gain under RA configuration are reported as below.

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

ClassB: -2.27%, -3.03%, -3.53%

ClassC: -2.90%, -3.14%, -1.04%

ClassD: -8.57%, -1.34%, 0.30%

Results for A classes was not available yet.

It was commented that enforcing NNSR in anchor for classes B, C and D might be wrong, as it is known to give losses. Therefore, the proposal might just show better results, because the adaptive approach may decide to never use NNSR in theses classes (and not better than NNVC without NNSR).

Further study encouraged with results for classes A. It is also suggested to study the amount of rate needed for network parameters, which should establish a higher percentage in low resolution classes.