JVET-Q0305 Disallowing JCCR mode for ACT coded CUs [H. Dou, L. Xu, Y.-J. Chiu (Intel)]
In 4:4:4 mode of Work Draft 7 (VTM7.1), ACT could convert a CU’s luma and chroma channel residuals into YCgCo colour space and then encode on the converted residual in YCgCo space. Followed by ACT on a CU, JCCR could be further applied to jointly encode the two converted chroma channel residuals in YCgCo colour space. The combination of enabling both ACT and JCCR within a CU increases undesirably the length of residual construction pipe at video decoder side. This contribution proposes not to enable JCCR for the ACT coded CUs to improve the worst-case residual reconstruction pipe delay at decoder side along with the benefit to reduce the RDO checks at encoder side. With HEVC SCC test clips, the experiments report the BD rate of the proposal has a very small impact in RGB test clips, and a very small improvement in YUV test clips (in LDB case).
- When dual-tree is off
- For RGB test clips, the average {G, B, R} BD-rates are {0.15%, 0.16%, 0.10%}, {0.25%, 0.20%, 0.09%} and {0.11%, 0.26%, -0.13%} for AI, RA and LDB, respectively
- For YUV test clips, the average {Y, U, V} BD-rates are {0.00%, 0.01%, 0.00%}, {0.02%, 0.02%, 0.03%} and {-0.02%, -0.03%, -0.04%} for AI, RA and LDB, respectively
- When dual-tree is on
- For RGB test clips, the average {G, B, R} BD-rates are {0.14%, 0.11%, 0.01%} and {0.02%, 0.25%, -0.09%} for RA and LDB, respectively
- For YUV test clips, the average {Y, U, V} BD-rates are {0.00%, 0.01%, -0.01%} and {0.00%, -0.11%, -0.10%} for RA and LDB, respectively
It is observed that for camera captured content in RA JCCR still seems to work with ACT YCoCg (loss is 0.4% when disabled. The proposal suggests a block-level change.
JVET-Q0506 proposes exactly the same, but provides additional data which show that the loss is much larger when JCCR is disabled at high level when ACT is used.
The complexity is not very detailed and does not provide enough evidence that a block level change is needed here. Furthermore, the loss of 0.4% in camera captured content (which is more critical to encode than screen content) is undesirable.
Further study of complexity criticality is suggested.
No action was taken on this.