JVET-Q0527 AHG9: Gradual Decoding Refresh for VVC [L. Wang, S. Hong, K. Panusopone (Nokia)] [late]

This contribution was discussed Wednesday 15 January at 1720 in Track A (chaired by GJS).

This contribution proposes a change of the gradual decoding refresh (GDR) capability of VVC, in which a GDR picture is separated by a virtual boundary into “clean area” and “dirty area”. It is proposed that, for CUs in the clean area, the virtual boundary is treated as picture boundary and all the rules associated with a picture boundary in VVC apply. For CUs in the dirty area, the virtual boundary is treated as no boundary. In addition, for CUs in the clean area, it is proposed that a separate HMVP table is maintained and updated with the candidates only coming from previously-coded CUs in clean area.

Currently, the encoder must constrain its mode selections in the current picture to avoid referencing samples in the “dirty area”. The encoder must also constrain inter prediction to avoid referring to the “dirty area” of the reference picture. And if the co-located block is in the “dirty area” of the reference picture, TMVP cannot be used.

The clean/dirty areas in the current picture and in the reference picture(s) may be different.

It is asserted that there is a coding efficiency impact from the current scheme and that it could be difficult to make an encoder that operates properly.

Currently we have a virtual boundary that can disable in-loop filtering across a boundary, but that boundary does not affect intra prediction or inter prediction referencing.

We also have subpictures, established at the sequence level, than can treat subpicture boundaries as picture boundaries.

For the proposed method, the following would apply when the proposed flag is equal to 1:

• Intra prediction would treat the virtual boundary position in the current picture as a picture boundary for purposes of referencing other samples of the current picture for intra prediction
• The decoder would store the virtual boundary position for each picture and would treat that boundary as a picture boundary when the picture is later referenced by other pictures.

The boundary granularity would be Max(8, MinCuSizeY). The benefit relative to the current scheme is reportedly about 3.3% (for classes A, B, C, D, E).

It was commented that this shifts the implementation burden from the encoder to the decoder, requiring additional granularity of processes in the decoder (a boundary in the middle of a CTU) and requiring extra testing effort.

There was discussion of potentially constraining the boundary to CTU granularity.

There was discussion of the granularity of the proposed processing and what was expressed in the proposed text. The proposed text and software had CTU granularity for the picture partitioning, but the simulation had used 8-sample granularity for motion compensation processing.

It was commented that there would be less of an implementation concern if the granularity was at the CTU level.

No action was taken on this.