JVET-AC0276 EE2-related: Complexity reduction for Decoder Side Control Point Motion Vector Refinement (test 2.3b) [M. Bestel, F. Le Léannec, P. Andrivon, M. Radosavljević (Xiaomi), H. Huang, Y. Zhang, Z. Zhang, C.-C. Chen, V. Seregin, M. Karczewicz (Qualcomm)] [late]
This contribution attempts to optimize the affine DMVR algorithm of test EE2-2.3 that was initially proposed in JVET-AB0178, in terms of trade-off between coding efficiency increase and complexity. It brings several aspects on top of test 2.3b, which are:
- Distortion based early termination
- 6-parameters affine model early termination
- Affine DMVR usage according to CU size and shape
- Adaptation of affine DMVR merge candidate list construction
The proposed modification of top of test EE2-2.3 reportedly leads to the following BD-rates and runtimes relative to ECM-7.0 anchor.
The proposed method reportedly leads to the following BD-rate performances over ECM-7.0 anchors.
RA : XXX% / -XXX% / -XXX% (Y/Cb/Cr). Runtime : XXX% (Enc), XXX% (Dec)
From the partial results shown, about 2/3 of the gain from 2.3b is retained, while the increase in encoder/decoder run time is reduced to about 1/3. From the current results, this would not be a reasonable tradeoff yet.
Further improvements are foreseen by proponents.
It was agreed to investigate this in an EE.
It was noted that another proposal that could be further investigated in a next EE might be 2.4c for which the results were not complete by the time of current EE review.