JVET-AD0169 AHG12: RPR filters for scale factors below 1.5x [J. Samuelsson-Allendes, S. Deshpande (Sharp)]

This input document relates to Reference Picture Resampling (RPR) in ECM. The document includes a proposal to add new Motion Compensation (MC) filters for downsampling with scale factors in the range from 1.1x to 1.35x. It is also proposed to modify the non-normative downsampling included in the ECM software for the same range of scale factors, asserted to improve the quality of the downsampled pictures. Combined, these changes result in an average BD-Rate difference (Y/U/V) of ‑5.70%/‑3.31%‑2.81% for 1.25x RA, ‑7.16%/‑2.42%‑2.00% for 1.25x LD, ‑9.30%/‑4.17%‑3.49% for 1.33x RA, and ‑11.65%/‑2.79%‑1.86% for 1.33x LD. In the first version of the proposal these average numbers do not include class A1 and A2 for RA, and the BQTerrace sequence is not included in class B for the 1.25x LD test case since the results are not available yet. In the second version of the proposal, the awaited results have been added (except for ParkRunning3 which is still awaited for 1.33x RA). Additionally, run time numbers were corrected and the RA results have been updated for class B, C and D since the class specific settings were not used in the results presented in the first version.

Gains reported above are for combining non-normative and normative part of the proposal under RPR testimg condition. The aspect of downsampling for prediction a lower resolution in the loop would be normative, whereas the downsampling before input to the encoder is non-normative. Results for only doing a non-normative change are not complete yet.

Current downsampling filters of VTM and ECM (10-tap filters are used fore 1.5x, which were inherited from SHVC) may be too aggressive in cases of downsampling ratios <1.5x.

The non-normative change is definitely beneficial, as it is relevant to have suitable filters that accommodate any downsampling ratios, even if not used in any CTC. This should be done in both ECM and VTM.

For the normative change, which could only be applied to ECM, it would be relevant to know its additional benefit, which can assessed when the delta relative to the non-normative approach is known. It was further requested to also report the so-called “PSNR1” for the normative aspect which measures the PSNR at coded resolution. In the results above, “PSNR2” is reported which is measuring before downsampling and after upsampling.

A follow-up report was given on Thursday 27 Apr. at 1240 (classes A not ready yet in ECM). The results indicate also relevant gains for PSNR1 in all sampling ratios, such that it can be concluded that the normative has some additional benefit.

Decision (SW): Adopt JVET-AD0169 (non-normative aspect) to ECM SW(+description?), and also VTM SW+description.

The normative aspect was agreed to be investigated in an EE