JVET-M0265 AHG16: Clean-up on MV Rounding [K. Zhang, L. Zhang, H. Liu, J. Xu, Y. Wang, P. Zhao, D. Hong (Bytedance)]
This was discussed in Track A Wednesday 16 January 1600-1630 (chaired by GJS).
In VTM-3.0, MV averaging is performed by pair-wise merge candidate, triangular prediction and affine prediction for chroma components with different ways of rounding. It is proposed to unify the MV rounding operation to be the same as MV rounding in affine MV derivation. In addition, it is proposed not to clip MVs for the luma component before calculating the MV for chroma components in affine prediction, to align the VTM software and the working draft. Simulation results reportedly show 0.00% BD-rate change under Random Access (RA) configurations.
Decision: Change the software to match the WD to remove clipping of luma MVs before deriving chroma MVs.
It was reported that prior to VTM 3, there was only one rounding in MV averages.
Offset = 1 << (F-1);
M= S >= 0 ? (S + Offset) >> F: -((-S + Offset) >> F).
This is rounding away from zero. Some recent changes to VTM 3 were not consistent with that, and it was proposed to always use this same rule.
Decision (BF/consistency): Adopt rounding away from zero for MV averages.