International organisation for standardisation

NSQT w/o 2x8 + JCTVC-F379 AMP (1.3%) – Decision: This combination was Adopted

Later discussion in Tue plenary: Does AMP deblock at non-8x8 boundaries? A: No, so it's OK.

Additional later remark: Might AMP be introducing some visual artifacts because of this? (e.g. 4x8, 4x16 artifacts). A proponent indicated that in their informal experience, they had not noticed an issue, although no formal subjective analysis had been conducted.

As a larger issue, not just in regard to AMP, it seems worth consideration of whether we are losing any significant quality by not deblocking at non-8x8 edges in general. Note that AVC has such deblocking.

Initially, the adoption was pending HM 3.3 integration and review and study of inferred merge RD incl. runtime impact during the meeting. The status was then reviewed Thu 8:45pm with no reported surprise, although the software needs some (but not an excessive degree of) work in regard to coding guidelines: needs work on compile errors, variable names, brackets, indentation, and memory allocation).

AMP only (enabled or disabled in CC)

The last option above using OBMC with just one-line overlap (being tested, roughly 2.1%)

NSQT (w/o 2x8) with AMP w OBMC (rough guess 2.0%)

AMP and NSQT provide more gain in LD (esp. Class E), roughly making up for less effective prediction by more effective residual coding.