JVET-AC0116 EE1-3.2: neural network-based intra prediction with learned mapping to VVC intra prediction modes [T. Dumas, F. Galpin, P. Bordes (InterDigital)]

This contribution reports the results of EE1-3.2.2. In EE1-3.2.2, VTM-11-NNVC with Filter-Set-1 activated, Filter-Set-0 deactivated, and the low-complexity version of the neural network-based intra prediction mode activated, all the neural networks being in 16-bit signed integer, must be run. This run aims at showing that the BD-rate gains of the low-complexity version of the neural network-based intra prediction mode inside VTM-11-NNVC are almost equivalent when this neural network-based intra prediction mode is combined with either Filter-Set-0 or Filter-Set-1.

It is reported that, on top of VTM-11-NNVC with Filter-Set-1 activated and Filter-Set-0 deactivated, the low-complexity version of the neural network-based intra prediction mode yields -3.21%, -3.52%, -3.38% and -1.54%, -1.01%, -1.25% of average BD-rate gains in AI and RA respectively.

On top of VTM-11-NNVC with Filter-Set-1 deactivated and Filter-Set-0 activated, the low-complexity version of the neural network-based intra prediction mode yields -3.35%, -3.60%, -3.65% and -1.57%, -1.08%, -1.20% of average DB-rate gains in AI and RA respectively.

This was presented in session 12.

Performance in combination with the filter sets #0 and #1 was almost identical.

Both training and inference cross-checks confirmed the results

Decision: Adopt JVET-AC0116.