JVET-AD0045 AHG10: Encoder MV selections and DMVR revisited [K. Andersson, R. Yu (Ericsson)]

In JVET-W0061 it was pointed out that DMVR can cause visual artefacts at subblock boundaries when it fails to find the correct motion and an encoder-only fix was proposed. The solution at that time was deemed to give too much objective overhead. In this contribution additional criterions to disable DMVR on the encoder side are proposed to reduce the overhead while removing visual artefacts. Firstly, the use of DMVR is always restricted for temporal layers lower than the highest temporal layer. When frame rate is equal or lower than 30Hz the use of DMVR is also restricted for the highest temporal layer. In higher temporal layers and when frame rate is higher than 30Hz it is generally better coherence between reference pictures which gives less problem to find the correct motion. If artefacts happen in the highest temporal layer they can not propagate to other pictures. Secondly, it is avoided to disable DMVR when average absolute subblock boundary differences are small when possible DMVR artefacts are less prominent.

The BDR effect on random access for VTM-20.0:

• SDR CTC RA Y/Cb/Cr: 0.04%/0.06%/0.05%
• HDR CTC RA DeltaE/PSNRL wY/wCb/wCr/: 0.07%/0.09% 0.08%/0.09%/0.13%

Similar encoding and decoding time as anchor.

This is an encoder only proposal that reportedly brings subjective quality improvement with some objective performance penalty, where the objective performance penalty has been reduced compared to the related previous contribution JVET-W0061.

One visual example in still picture form is provided in the PPT file attachment of this proposal.

Cross checker reported that the objective results are matched, and that visual check was performed. It was reported that most of the videos looked similar, and one specific case of visual quality improvement was found and included in the crosscheck report.

It is suggested to confirm the reported subjective quality improvement in video mode.

It was asked if the proposed method has been tried on BDOF or on ECM (which has a more complicated DMVR process). It was reported that no visual artefacts associated with BDOF have been observed, and that this has not been tried on ECM.

The proponent does not suggest to enable the method in the CTC, but suggest to enable the method by default when GOP-based RPR (which is already supported by VTM) is enabled.

It is suggested to organize an informal viewing session to confirm the visual benefits.

An informal viewing session was conducted on Thursday 27 April, a report of this was given during the joint meeting with AG 5. 6 people had participated. It was confirmed that there is visual benefit.

Decision (SW): Adopt JVET-AD0045. No impact on CTC, but requires update of VTM description