JVET-AE0288 Results of expert viewing for the spatial scalability category of the VVC multilayer VT [M. Wien (RWTH Aachen University), K. Andersson (Ericsson), P. de Lagrange (InterDigital)]
This document reports the results of an on-site expert viewing performed at the 31st JVET meeting in Geneva, CH.
The evaluation targets at the spatial scalability category of the test plan document JVET-AD2021, comparing the compression performance of dual layer (DL) coding with a downscaled base layer at scaling ratio 2 to single layer coding at the low resolution with upscaling to the target resolution using the VVC interpolation filters (UP) as well as single layer coding at the target resolution (SL).
A total of 29 experts participated in the expert viewing. Based on this larger database, the results reveal small confidence intervals and can reportedly be considered quite reliable.
With respect to the goals of this experiment, the following observations were made:
- DL shows a slight benefit over UP for DrivingPOV3 (overlapping confidence intervals), very minor for MountainBay2.
- No benefit or slight loss for DL compared to UP for Procession.
- More loss for DL compared to UP for CorsaireJoy, Marathon2, and, partially significant, for TallBuilding2.
- Relation of SL, DL, UP somewhat according to expectations on MountainBay2 (SL gaining at high rates, while DL, UP gaining at low rates.
- ISSUE observed for TallBuildings2: UP better than SL and DL for lower two rate points. Possible reason: less pumping artefacts for UP streams which are coded at a lower QP. Better detail of SL or DL does not compensate for this issue
- ISSUE observed for CorsaireJoy: UP rated better, although more detail is present in DL version. Inspection indicates that DL version also shows partially more visible artefacts.
The observed issues appear to be related to suboptimum bit allocation in the SL and DL configurations. It is recommended to take these results into account when setting up the set of four rate points for a formal evaluation.
It was asked whether the bit rates might be too low for demonstrating the benefit of coding at full resolution (both for SL and dual layer).
It was commented that potentially a DCR (degradation category rating) test is not most suitable to see the benefit of higher detail.
It was suggested to potentially look at intra frames in terms of quality and bit rate allocation between the different cases.
It was asked whether it would be possible to combine this analysis with test results from Antalya to conclude whether there are more cases where the upsampled low-res is consistently better than single layer.
Are the rate points appropriately selected? Does it make sense to go to very low quality with scalable coding? Is it possible to find a perform tests over a wide range of rates/qualities at same full resolution and get consistent results?
Statistics about rate allocation to base/enhancement should also be taken into consideration.
An informal viewing session was planned for Wednesday for interested participants.
AHG4: Test material (1)
Contributions in this area were discussed at 1600–1610 on Sunday 16 July 2023 (chaired by JRO).