JVET-N0030 CE10: Summary Report on Combined Inter and Intra Prediction [C.-W. Hsu, M. Winken]
This report was discussed in Track B on Thursday 21 March 0900 (GJS)
A summary of Core Experiment 10 (CE10) on combined inter and intra prediction is reported. One sub CE was created to investigate combined inter and intra prediction techniques on top of VTM-4.0, as follows:
- CE10-1: Combined inter and intra prediction (CIIP) with weight modifications.
There are 2 tests for the sub CE. All tests are evaluated based on the common test conditions defined in JVET-M1010. All tests and crosscheck results are integrated in this report.
Proposal Document # | Corresponding Tests | Title | Author(s) |
CE10-1.1 | CE10: CIIP with position-independent weights (Test CE10-1.1) | L. Pham Van, G. Van der Auwera, A. K. Ramasubramonian,V. Seregin, M. Karczewicz (Qualcomm) | |
CE10-1.2 | CE10: CIIP using explicit signalling of weights (CE10-1.2) |
# | Proposal | Supported intra modes | Weight scheme | Weight sets (wIntra, wInter) |
VTM-4.0 | 4: DC, PL, VER, HOR | PL, DC: equal weights | (4, 4) | |
VER, HOR: position dependent weights, four regions | (6, 2), (5,3), (3,5), (2, 6) | |||
CE10-1.1 | 1: PL | Implicit weight assignment | 2: (3, 1) 1: (2, 2) 0: (1, 3) | |
(number of intra-coded left and top neighbours) | ||||
CE10-1.2 | 1: PL | Explicit weight signalling | 0: (4, 4) 10: (2, 6) 11: (6, 2) | |
# | Proposal | Config. | VTM | ||||
Y | U | V | EncT | DecT | |||
CE10-1.1 | RA | -0.05% | -0.02% | 0.02% | 98% | 100% | |
LB | -0.03% | 0.20% | 0.07% | 98% | 100% | ||
CE10-1.2 | RA | -0.04% | 0.04% | 0.05% | 100% | 99% | |
LB | -0.01% | 0.16% | -0.01% | 101% | 98% |
The proposals are both substantially simpler than the current design (removing position-dependent weighting and an MPM list derivation). It was commented that it may also be desirable to consider a pure simplification of just removing the DC, horizontal and vertical modes. Estimated results for that are ~0.01% loss in RA and ~0.1% loss in LB with equal weights. Hypothetically, having weights determined by RA vs LB would reduce the loss in LB for that to about 0.03%.
It was commented that the explicit weight method of CE10-1.2 would enable additional gain if the encoder did more searching. The estimated benefit of that was about 0.09% overall and 0.17% in Class A1 for 3% encoder runtime (for a straightforward approach). The CE10-1.2 test did not use more full RD tests than the VTM anchor, but performed 12 SATD checks, whereas the CE10-1.1 scheme performed 4 SATD checks (where the anchor used 7 SATD checks, so one scheme decreased encoder runtime and the other one slightly reduced encoder runtime). The difference in decoder complexity between the two proposed methods did not seem substantial.
For the explicit signalling method, the first flag indicating whether a non-equal weight is bypass coded and the second flag is context coded. It was commented that the bypass coding of the first bin means that the encoder would be forced to check both possibilities to avoid wasting signalling bits. It was then commented that there could be some flag to turn off the low-level signalling. Another discussed possibility was to use a context for coding the first bin as well as the second one.
It was commented that since the encoder resources are higher for the CE10-1.2 scheme and the additional gain available for an even better encoder is quite limited, the CE10-1.1 scheme seemed preferable.
Decision (complexity reduction/cleanup): Adopt CE10-1.1 scheme.