Since DIF (used in JCTVC-D325) is no longer planned to be included in the HM, it was suggested to focus on JCTVC-D394.
Relative to the anchor, JCTVC-D394 reportedly provided approximately 1.3% gain, with roughly 7% increase in encoder complexity, and 5% increase in decoder runtime for RA and 5% decrease in decoder runtime for LD. This AMVR proposal was reportedly somewhat different than what had previously been in the TMuC. As proposed, in the HE case, the anchor uses 12 taps, and the AMVR proposal uses (the same) 12 taps for the 1/4 grid positions and 8 taps for the 1/8 grid positions.
On a PU basis, either 1/4 grid or 1/8 grid resolution MVs are indicated (using a flag). For bipredictive cases, each MV has (the equivalent of) such a flag.
It was noted that making this decision at the encoder side adds complexity. This complexity was mitigated in the proposal by the use of fast (hierarchical) search.
For chroma, 1/16th position bilinear MC interpolation was used for the proposal.
Since we have decided to change the luma and chroma interpolation filters, the benefit for this technique may be different in the modified context. The exact impact of that change is unknown.
It was remarked that the prior implementation of AMVR in the TMuC context seemed to have a messy impact on the software.
Some participants felt that the need to support the additional positions in the decoder may not be justified by the quantity of reported gain.
It was recommended to further test the scheme in the new context (e.g., in a CE).
Dostları ilə paylaş: | 
