JVET-Q0801 CE2/CE3-related: Lossless coding for palette coding [S. Yoo, J. Choi, J. Lim, S. Kim (LGE)] [late]
Presented in Track B Wednesday 15 January, 1630
In the current VVC draft text, Exponential-Golomb with 3rd order (EG3) is utilized for the escape value coding. However, the distribution of the maximum escape value varies significantly over QP values so that it could produce the number of bits more than an input bitdepth of a sequence. Moreover, at Brusssels meeting, EG5 is adopted for escape value binarization method. Therefore, a binarization selection method between the fixed length coding (FLC) and EG5 is proposed, by sending a high level flag to indicate which binarization to be used. For FLC binarization, the input bit depth (cMax for bin length) is not transmitted but derived using the syntax element, min_qp_prime_ts_minus4. The coding efficiency reportedly shows like below:
- Lossless
- TGM 1080p
- Dual Tree = 1 : -3.11% (AI) / -2.85% (RA)
- Dual Tree = 0 : -4.94% (AI) / -3.54% (RA)
- Overall
- Dual Tree = 1 : -1.49% (AI) / -0.99% (RA)
- Dual Tree = 0 : -1.95% (AI) / -1.10% (RA)
- TGM 1080p
- Low QP (2 ,7, 12, 17)
- TGM 1080p
- Dual Tree = 1 : -0.99% (AI) / -0.93% (RA)
- Dual Tree = 0 : -2.14% (AI) / -1.65% (RA)
- Overall
- Dual Tree = 1 : -0.55% (AI) / -0.51% (RA)
- Dual Tree = 0 : -0.97% (AI) / -0.70% (RA)
- TGM 1080p
For QP 2..12 and lossless, fixed length coding is used, otherwise EG5.
No support was expressed by other experts
No action was taken on this.