JVET-B0060 Improvements on adaptive loop filter [M. Karczewicz, L. Zhang, W.-J. Chien (Qualcomm)] [late]

Chaired by J. Boyce.

In this contribution, several changes to the adaptive loop filter (ALF) in HM16.6 JEM-1.0 are proposed. Three main introduced modifications are: classification with the diagonal gradients taken into consideration, geometric transformations of filter coefficients, and prediction from fixed filters. In addition, some cleanups of the software are also included. With the proposed methods, the coding performance of HM16.6 JEM-1.0 is reportedly improved by 1.1% and 1.1% on average for all intra (AI) and random access (RA) configurations, respectively, when all the tools are enabled and by 1.5% and 1.5% on average for AI and RA configurations, respectively, when only ALF is enabled. The overall performance improvement of ALF, compared to HM16.6, reportedly reaches 4.0% and 6.0%, on average for AI and RA, respectively.

The HM 3.0 version of the ALF was put into the JEM at an earlier stage. Changes are proposed with respect to the HM 3.0 version of ALF. It was commented that later versions of the HM (HM 8) had made changes to ALF vs. the HM 3.0 version.

The contribution proposes 25 classes rather than the 15 classes in the older design.

The proposal avoids use of temporal prediction in I frames.

Software cleanups were also proposed.

Simulation results were only provided for the combination of changes. Some of the changes are inherently grouped together, but some could be separated. It was specifically requested to provide test results with the fixed filters on and off. The impact of the change to the chroma filter shape alignment with luma filter shape is also requested.

Decision (SW): Make SW available in experimental SVN branch of JEM software. Include in the EE document for testing.