JVET-AF0059 AHG12: Fix to interpolation filter for intra prediction [K. Andersson (Ericsson)]

It was noted that ECM-10.0 can give stripe like visual artifacts and it was discovered that the reason for that was in interpolation in intra prediction due to non-unity gain of one of the 64-phase interpolation filters. It is therefore suggested to correct the filter with non-unity gain using a mirror version of corresponding mirror phase filter.

Objective performance (BDR) versus ECM-10.0 is reported as:

AI: Y: -0.01%, U 0.01%, V -0.05%

It was asserted that fixing the gain of the interpolation filter with non-unity gain also can improve subjective quality.

In ECM-10.0, the intra prediction interpolation at phase 41/64 has filter coefficients { 3, -21, 106, 180, -25, 3 } that sum up to 246 rather than 256. The proposal suggests to use the mirroring coefficients of phase 23/64 for phase 41/64, with coefficients { 3, -20, 106, 189, -25, 3}.

It was commented that this is the only phase that has this unintended behaviour.

It was also shown that fixing the coefficients brings subjective quality improvement.

Decision: Adopt JVET-AF0059.