JVET-Q0504 CE2-related: Palette mode for non 4:4:4 colour format [R.-L. Liao, M. Sarwer, J. Chen, Y. Ye, J. Luo (Alibaba), Y.-H. Chao, C.-H. Hung, W.-J. Chien, V. Seregin, M. Karczewicz (Qualcomm)]
This contribution proposes to enable the palette mode for non 4:4:4 colour format including 4:2:0, 4:2:2 and monochrome colour formats. To support the palette mode for non 4:4:4 colour format, two modifications are made to VVC draft 7. Firstly, for a sample only containing luma component, only the luma value is signalled if the sample is coded using escape mode. This is the same as in the HEVC SCC extension. Secondly, for a local dual tree block, a default value, defined as (1 << (component bit depth - 1)), is set to the new palette entry before the palette predictor update. It is reported that, for class TGM in 4:2:0 colour format, the overall performance impact for {Y, U, V} is {-6.35%, -6.21%, -5.88%} for AI, {-3.96%, -3.68%, -3.53%} for RA and {-1.42%, -1.28%, -1.27%} for LB.
The approach is similar as in HEVC. Tests are with dual tree on
For class F, the luma gains are 1.24%/1.35%/0.64% in AI/RA/LB.gains for dual tree on in case of class TGM with 4:4:4 palette were in the range of 16%/11%/5% (AI/RA/LB) for AI
The gains for 4:2:0 are somewhat still in the same range (or even lower) as they were several meeting cycles ago (in March 2019, it was reported that palette gain for AI in TGM 4:2:0 could have around 8% gain). Since then, other competing tools such as BDPCM and TS have been included.
There is no evidence why the decision made earlier (not including palette for possible 4:2:0 profile) should be revised.
The proposed approach is a minimum change to the palette design of VVC draft 7, allowing the combination of palette with local dual tree (such that it would be runnable with 4:2:0 content).
When adopting such a change of the palette tool, this would not mean to include palette in a 4:2:0 profile. It would just enable a 4:4:4 decoder that receives a 4:2:0 or 4:2:2 or 4:0:0 stream using palette tools to decode that stream (where the stream should be compliant with a 4:4:4 profile which includes palette).
Decision: Adopt JVET-Q0504.
In a follow-up discussion in the plenary of Sunday 12 January (see the notes in section 11.2), it was decided not to include the palette mode in the current 4:2:0 profile. It was further pointed out that the high-level enabling flag needs to be sent independently of chroma format. The profile constraint needs to be implemented.