JVET-AO0191 [AHG11]: DCVC-RT as an End-to-End Learned Intra Frame Codec [M. Aderdor, T. Solovyev, E. Alshina (Huawei)]

This contribution reports the coding gains when using DCVC-RT as an I-frame coder on top of VTM 23.13. DCVC-RT is the next iteration in the DCVC series after DCVC-FM. The external decoder has a complexity of 365 kMACs/pixel and 36.4 M Parameters. The coding gains are as follows:

On top of NNVC 15.1 configured as VTM 23.13:

AI: -8.4% (Y) -16.1% (U) -15.4% (V)

RA: -1.7% (Y) -8.4% (U) -5.4% (V)

On top of NNVC 15.1 configured as VTM 23.13 + DCVC-FM I-frames:

AI: -3.3% (Y) -5.3% (U) -5.2% (V)

RA: -0.7% (Y) -2.7% (U) -2.1% (V)

It was concluded to include this in EE1 as replacement for DCVC-FM.