JVET-P0353 CE1-related: Reference picture resampling filters [J. Luo, Y. Ye, M. Sarwer (Alibaba)]
CE1 test 3 in JVET-O2021 tests the application of a 12-tap SHM filter for reference picture down-sampling. In order to avoid an increase of complexity and memory access bandwidth, this contribution proposes to use cosine windowed sinc filters with the filter length, gain and number of phases matching those of the existing motion compensation interpolation filters (MCIF) in VVC. The simulation results reportedly show that comparing to the CE1 anchors which use existing MCIF for reference down-sampling, average BD rate deltas of PSNR1 are { -5.96%, -2.44%} for ratio 2:1 and 1.5:1, and average BD rate savings of PSNR2 are { -3.80%, -1.78%} for ratio 2:1 and 1.5:1.
In the CE, the filter for affine was not modified. In this proposal, the affine filter was modified. This probably doesn't make much of a difference, and the proponent was running tests to confirm that.
It was commented that the sequences that tend to benefit the most from affine are, at least mostly, not tested here.
The half pel position is not using the proposed filter. The proponent was running tests to confirm the effect of that. The same issue existed in the CE.
Compared to the CE1-1.1 result, for the 2x case, there is a reported benefit of about 1.5% for PSNR1 LB.
It was commented that the filter coefficients in the first and second versions were different.
There was discussion of whether it is a problem that the filter has a gain higher than 1 for some frequencies. Some participants indicated that since RPR is not applied in a cascaded manner, this should not be a problem.
A participant suggested that, regardless of what action is taken at this meeting, further study be conducted, perhaps including subjective viewing and alternative test sequences.
The SHVC filter also has a "bump".
Another participant commented that HDR testing may be desirable.
Focusing on the 2:1 case, it was commented that this filter has less aliasing than the one from the CE. The JVET-P0353 filter was submitted quite late, although the difference is only the values of the taps. Some tests for JVET-P0353 were still ongoing. The actual difference for the switching pictures is larger than what was measured here, and the difference was significant.
Decision: JVET-P0353-v5 2:1 8-tap & 4-tap filter, subject to checking for any issue in ongoing affine and half-pel tests.
For 1.5:1, stay with the CE1 result.