JVET-B0058 Modification of merge candidate derivation [W.-J. Chien, J. Chen, S. Lee, M. Karczewicz (Qualcomm)] [late]

Chaired by J. Boyce.

In this contribution, modifications to the merge candidate derivation are proposed, including higher motion resolution, POC-based merge candidate pruning, simplification of the advanced temporal motion vector predictor, and derivation of the spatio-temporal motion vector predictor. With the proposed modifications, an additional 0.6%, 0.6%, and 2.3% BD rate reduction over HM16.6 is reportedly achieved for random access, low delay B, and low delay P configurations, respectively.

Four different modifications are proposed, but only combined performance results are available, not individual tool performance numbers.

It was proposed to move from 1/8 pel to 1/16 pel storage and merge candidate accuracy. No additional storage or memory bandwidth vs. HEVC is used, but the allowable range is decreased to ~4k.

This was further discussed Tuesday AM (chaired by JO/JB). It was reported that the higher precision of MV storage increases by 0.0/0.3/0.4% for the cases of RA/LDB/LDP.

It was asked if the gain will be reduced once the harmonization of the affine with OBMC is included in the JEM. The gain of this aspect is greater than the affine OBMC harmonization.

Decision: Adopt the 1/16 pel motion vector storage accuracy to the JEM and the common test conditions.

It was suggested for individual “tool on” experiments to only test this with ATMVP turned on, because these would show more benefits together.

Decision (SW): Add to the experimental branch the pruning and ATMVP simplification. Include in the EE document for testing. Separate experimental results for each proposed aspect should be provided.