JVET-Q0291 CE2-related: On maximum palette size of VVC [M. G. Sarwer, Y. Ye, J. Luo, R. -L. Liao (Alibaba)]
In VVC7, the maximum predictor size for both joint and separate palettes are set to 63 and the maximum palette size for both joint and separate palettes are set to 31. However, the complexity of the palette generation and predictor update of separate tree is roughly double that of single palette since the separate tree need to process two palette tables for luma and chroma. This contribution proposes to reduce the maximum predictor and palette size of separate palette to the half of that of joint palette.
Following results are reported as compared to VTM-7.0.
- YUV 4:4:4 (dual tree ON):
- Overall: -0.03% (AI), 0.05% (RA), 0.01% (LB) BD-rate saving
It is also proposed to signal the parameters for max palette size and max predictor size in SPS (as is the case in HEVC).
Question: Is there really a decoder complexity problem? In dual tree, probably operations can be performed in parallel. Nevertheless, the processing power for the palette prediction and table generation is doubled.
It is also mentioned that palette is more efficient in single tree case; the loss is marginal.
The suggested change of reducing the palette size to 15 and predictor size to 31 for dual tree appears reasonable. This aspect was agreed for adoption.
For the aspect of table size signalling, see further notes under JVET-Q0519.
Decision: Adopt JVET-Q0291, only the aspect that the palette size is reduced to 15 and the predictor size to 31 for the dual tree case.