JVET-R0380 Scaling list for adaptive colour transform [S. Iwamura, S. Nemoto, A. Ichigaya (NHK), K. Naser, P. de Lagrange, F. Le Leannec, P. Bordes (InterDigital)]

In this contribution, modifications of scaling matrix derivation are proposed. In VVC Draft 8, adaptive colour transform (ACT) for 4:4:4 can be controlled at the CU level flag (cu_act_enabled_flag). Although the energy distribution of transform coefficients in RGB and YCgCo domain would be different, in VVC draft 8, user defined scaling lists are always shared for G and Y, for B and Cg, for R and Co, which may cause undesirable behaviour of quantization process. To avoid the undesirable behaviour, this contribution proposes to introduce additional flag (scaling_matrix_act_disabled_flag) in APS to disable scaling matrix for the block coded with ACT.

V2 document add solution #2 which repurposes the existing “scaling_matrix_for_lfnst_disabled_flag” to disable scaling list for both LFNST and ACT coded block.

V3 document add solution #3 and #4, which add a new SPS flag instead of APS.

The initial solution and solution 2 are obsolete due to the adoption of scaling list signalling at SPS in JVET-R0064.

Solution 3 is proposing a similar (but separate) mechanism for ACT

Solution 4 is proposing using the same flag that disables usage of scaling list for LFNST also for ACT.

There were different opinions about the necessity of disabling scaling matrices for ACT on.

It was also asked if the three transforms are used in combination? They probably are, but this does not seem to be a problem.

There is the general opinion that such a mechanism would be useful, as the charcteristics of RGB and YCoCg in the spatial domain are probably different. Solution 3 seems to be more appropriate, as the characteristics of LFNST and ACT are probably different, and the additional flag would only be relevant in a 4:4:4 profile.

It is asked if it is more appropriate to switch off the scaling matrices for RGB coded blocks rather than YCoCg? Would it be better to design the scaling matrices for the latter and disable for the RGB coded blocks? It was also asked how it is handled in HEVC, and whether there are more blocks coded in YCoCg or in RGB typically.

Was further reviewed Wed. 22 0640 UTC. Numbers of frequency of usage for RGB/YCoCg are presented. As a general tendency, YCoCg is more frequently used in high QP, and RGB is more frequently used in lower QP. Three are two sequences where YCC is preferred over the entire CTC QP range. If however blocks in skip mode are not counted, it is more evident that a sequence dependency exists.

Another new solution (called solution 3-3 in the new version zip v4) is proposed that allows switching, either disabling scaling matrices for RGB or YCC. This is only existing in the slide deck

Decision: Adopt JVET-R0380 solution 3-3. Proponents shall provide the spec text in a new upload. It is to the discretion of the editors to align the naming of new syntax elements in a consistent way (e.g. “…designate_rgb_flag” does not appear appropriate)

AHG11: Screen content coding (9)

Initially presented in AHG sessions 2.3 Thu 9 April 1300-1500 and 2.5 Tue 14 April 0700-0900 (chaired by JRO)