JVET-AM0214 AHG 11: Motion Vector Restriction in Multi-Layer Hybrid NN-based and Conventional Coding [F. Brand, T. Solovyev, E. Alshina (Huawei)]
In the April meeting, JVET decided to start developing a framework for including NN-coded reference frames into VVC / NNVC. In an approach, multi-layer coding is used, where the base layer consists of NN-coded frames while the enhancement layer consists of VTM coded frame, which have the NN-coded frames as reference. This proposal shows that omitting the motion vector search and transmission, and instead setting them to zero gives a encoding runtime reduction, without noticeable performance change.
AI -0.02%/0.25%/0.20% 88% EncT
Loss in chroma is only coming from one sequence (Basketball). The reason for that seems to require further
Only merge mode is kept.
It is agreed that it definitely makes sense to disable motion estimation in the context of the multi.layer interface for NN intra base layer, also as for more local adaptivity the blocks in base and enhancement need to be co-located.
It is commented that the effect of encoder runtime reduction would be less relevant for RA.