JVET-V0177 CE-related: Additional information for CE on HBD coding [A. Browne, T. Hashimoto, H. Jhu, D. Rusanovskyy, K. Kawamura, T. Zhou] [late]
This was presented and discussed in session 20, 0010-0130 UTC, Tuesday 27 April (chaired by JRO and GJS)
This contribution presents additional results on the RRC methods for high bit depth video coding being studied in CE. Results include performance comparison for selected RRC methods against the HM16.23. as well as additional results on the performance of the Method 2 presented in JVET-T0105 [6]. Complexity increase compare to VVC is provided for selected RRC methods.
Results below have been produced by comparing reported results for CE tests against the HM16.23 anchor, provided by AhG8. Complete Excel tables are provided along this contribution.
The table below shows coding gain achieved by selected methods against the HM16.23. Impact of EGk invoking parameters change on top of selected tests is provided for SVT16 data separately in the CE3.1 (EGk) column.
Summary of results (BD-rate gain) for selected RRC tests, CE CTC, LowQP test configuration.
HBD CTC | CE3.1 (Egk) | |||||||||
Test | HDR PQ | HDR HLG | SVT12 | SVT16 | SVT16 | |||||
wY | wU | wV | Y | U | V | Aver. | Aver. | Aver. | ||
AI | VTM12 | -6.79% | -9.19% | -9.75% | -3.68% | -6.06% | -6.19% | 0.46% | 34.93% | 26.67% |
AnchorCE | -7.24% | -9.64% | -10.23% | -4.09% | -6.39% | -6.53% | -2.17% | 0.77% | 0.24% | |
CE1.4-A | -7.56% | -9.85% | -10.44% | -4.57% | -6.67% | -6.82% | -3.56% | -1.55% | -1.74% | |
LDB | VTM12 | -6.61% | -7.67% | -8.77% | -5.14% | -8.18% | -7.01% | -1.35% | 21.09% | 17.88% |
AnchorCE | -6.68% | -7.73% | -8.82% | -5.12% | -8.20% | -7.06% | -1.79% | 0.58% | 0.35% | |
CE1.4-A | -6.73% | -7.79% | -8.89% | -5.14% | -8.21% | -7.05% | -2.43% | -0.99% | -1.08% | |
RA | VTM12 | -6.81% | -7.91% | -8.64% | -5.72% | -8.64% | -7.32% | -1.64% | 19.55% | 16.42% |
AnchorCE | -6.93% | -8.04% | -8.77% | -5.73% | -8.68% | -7.37% | -2.11% | 0.32% | 0.08% | |
CE1.4-A | -7.01% | -8.10% | -8.85% | -5.77% | -8.69% | -7.38% | -2.67% | -1.14% | -1.22% | |
The table below shows coding gain achieved by enabling extended precission flag, as per CE3.1, for selected methods against the HM16.23 for 12 bits data.
Summary of results (BD-rate gain) for selected RRC tests, CE CTC, LowQP test configuration.
CE3.1 Extended precission flag + Test vs. HM16.23 | ||||||||
Testvs HM | HDR PQ | HDR HLG | SVT12 RGB | |||||
wY | wU | wV | Y | U | V | Aver.GBR | ||
AI | VTM12+ce3.1 | -7.88% | -10.64% | -11.29% | -4.16% | -6.81% | -6.96% | -0.41% |
AnchorCE+ce3.1 | -8.35% | -11.13% | -11.81% | -4.54% | -7.14% | -7.30% | -3.27% | |
CE1.4-A+ce3.1 | -8.63% | -11.31% | -11.99% | -5.00% | -7.38% | -7.54% | -4.06% | |
LDB | VTM12+ce3.1 | -7.21% | -8.24% | -9.37% | -5.69% | -8.85% | -7.78% | -2.07% |
AnchorCE+ce3.1 | -7.27% | -8.30% | -9.44% | -5.69% | -8.88% | -7.81% | -2.48% | |
CE1.4-A+ce3.1 | -7.31% | -8.35% | -9.49% | -5.71% | -8.88% | -7.79% | -2.86% | |
RA | VTM12+ce3.1 | -7.32% | -8.44% | -9.22% | -6.17% | -9.25% | -8.01% | -2.30% |
AnchorCE+ce3.1 | -7.45% | -8.58% | -9.37% | -6.19% | -9.29% | -8.05% | -2.74% | |
CE1.4-A+ce3.1 | -7.52% | -8.64% | -9.44% | -6.22% | -9.29% | -8.05% | -3.07% | |
Summary of results (BD-rate gain) for selected RRC tests, CE CTC, lossless test configuration.
Test | bit-rate saving | ||||
HDR PQ | HDR | SVT12 | SVT16 | ||
AI | VTM | 2.93% | 7.05% | 31.02% | 73.47% |
AnchorCE | 0.36% | 2.82% | 3.12% | 1.46% | |
CE1.4-A | -1.97% | -0.77% | 0.15% | 0.05% | |
LDB | VTM | 3.67% | 2.09% | 29.88% | 80.20% |
AnchorCE | 4.10% | 3.96% | 2.71% | 1.18% | |
CE1.4-A | 0.90% | -0.52% | 1.16% | 0.79% | |
RA | VTM | 3.30% | 1.85% | 29.30% | 80.25% |
AnchorCE | 3.80% | 3.79% | 2.55% | 1.11% | |
CE1.4-A | 0.70% | -0.59% | 1.10% | 0.74% | |
Benefit of history based vs. local based is around 0.3% for AI, 0.1% for RA in case of 12 bit low QP
Gain over HM is around 7% for 12 bit low QP in PQ, less for HLG
Additional gain of extended prec is >1% for AI, 0.5% for RA
For 16 bit data, only history based methods are able to provide (small) gain over HM
Definitely a modified entropy coding method (and extended transform precision) is necessary for professional application scenario, in particular 16 bit case.
Profile definitions have not yet been discussed, one aspect that might considered if there will be different 12 bit profiles (or tiers) for consumer and professional use cases.
Is the technology investigated in the CE mature for being adopted?
For TSRC, the method from the CE anchor (CE2.1) should be adpted, as it is simple, and the more complex solutions do not provide significant gain.
Decision: Adopt JVET-V0054, method CE2.1 as entropy coding method for high bit depth in TSRC
For RRC, two major candidates, CE1.4-A, and JVET-V0084 (which is a CE related contribution, modification of CE1.6)
In CE1.4-A, the history based approach can be disabled (then it falls back to 1.2). An alternative is proposed in JVET-V0106 (CE related), where the same history based method of CE1.4-A I built on top of CE1.1 (the latter almost identical with the CE anchor. CE1.1 is more hardware implementation friendly than CE1.2. All these have been thoroughly investigated in CE
JVET-V0084 is only history based (cannot be disabled)
Decision: Adopt JVET-V0106 as entropy coding method for high bit depth in RRC. High level flag to disable the history based part.
Decision: Adopt JVET-V0047 CE3.1 method for the high precision computation of transform scaling