JVET-AD0180 Non-EE2: Affine AMVP mode with one MVD [H. Huang, V. Seregin, Z. Zhang, M. Karczewicz (Qualcomm)]

This contribution presents an affine AMVP mode that only signalled one MVD. The MVD is added to all the CPMVs of an affine MVP to generate the final CPMVs. The simulation on top of ECM-8.0 reports {-0.09%(Y), -0.07(U)%, -0.06%(V)} BD-rate change in RA, and {-0.09%(Y), 0.02%(U), -0.05%(V)} in LDB.

Was it also tried for 4-parameter affine? Yes, but benefit would be small.

Encoder runtime increase with up to 9% in the report, which may not be reliable, but definitely some additional encoder checking necessary. This would not be an acceptable tradeoff. After completion of cross-check, the increase of encoding time was confirmed, Further, the proponents showed preliminary results (only classes C and D) which had an increase in encoder runtime by only 1%, but the gain was only 0.05% due to omission of some encoder checks.

It was agreed to investigate this in an EE, with the goal of keeping more of the original gain with acceptable encoder runtime.