JVET-Q0098 On QP Adjustment for Adaptive Colour Transform [T. Tsukuba, M. Ikeda, Y. Yagasaki, T. Suzuki (Sony)]
In VTM-7.1, after applying QP adjustment (-5, -5, -3) for ACT, the quantization parameter qP for TS can be less than QpPrimeTsMin, and the qP for non-TS can be less than 0. To keep the qP in the valid range, this contribution proposes to a) apply the QP adjustment for ACT before clipping the qP for TS to QpPrimeTsMin, and b) clip the qP for non-TS to 0 after the QP adjustment for ACT.
For DualITreeOff at low QP condition (QP=2, 7, 12, 17), it is reported that BDrate-differences of (G, B, R) are (-0.17%, -0.38%, -0.33%) for AI, (-0.09%, -0.25%, -0.27%) for RA and (-0.11%, -0.34%, -0.31%) for LB; BDrate-differences of (Y, U, V) are (0.00%, 0.00%, 0.00%) for AI, (0.00%, 0.00%, -0.01%) for RA and (-0.02%, 0.01%, -0.01%) for LB
For DualITreeOn at low QP condition, it is reported that BDrate-differences of (G, B, R) are (0.00%, 0.00%, 0.00%) for AI, (-0.06%, -0.25%, -0.29%) for RA and (-0.09%, -0.29%, -0.27%) for LB; BDrate-differences of (Y, U, V) are (0.00%, 0.00%, 0.00%) for AI, (0.00%, -0.05%, -0.07%) for RA and (-0.05%, -0.19%, -0.13%) for LB.
It is further pointed out by the proposnents that current VTM SW crashes when internal bit depth = 8. This however only happens when QP was out of range – not relevant after bug fix.
The proposal should resolve the issue without taking influence on deblocking in terms of the spec text. Cross-checkers however point out that the SW implementation seems to change deblocking via modified QP’.
Related:
- JVET-Q0142 (same solution for non-TS)
- JVET-Q0166 (somewhat different)
- JVET-Q0241 (same solution, but different results)
- JVET-Q0510/JVET-Q0511 (similar solution, but more elements such as signalling)