JVET-Q0761 BoG report of CE4 related contributions on inter prediction with geometric partitioning [H. Yang]

Review of this BoG report was discussed Saturday 11 January 1115 in Track A (chaired by GJS).

Four sessions had been organized for the BoG meeting, 27 contributions had been reviewed.

• Jan. 8, 2000 to 2230, on the category of mode reduction for geometric partitioning.
• Jan. 9, 0800 to 1300, on remaining contributions.
• Jan. 9, 1430 to 1725, on remaining contributions.
• Jan. 10, 1530 to 1730, on the summary table of various proposals in the category of mode reduction for geometric partitioning.

JVET-Q0061 had been discussed in Track A when the CE4 report was presented.

BoG-recommended normative changes, relative to the CE4 method:

• JVET-Q0188: change the sampling position from (4i+1.5, 4j+1.5) to (4i+2, 4j+2), for the calculation of motionIdx for the motion storage of GEO mode (method 2 of proposal).

With the proposed method, the weightIdx derived for blending matrix can be reused (if desired) for the derivation of motionIdx, and no need for separate calculation of motionIdx.

This is a small adjustment of the formula. It was commented that this might have no real reduction of complexity for most implementations (since they might not reuse the blending mask to derive the motion mask); however, it makes equations in different places consistent with each other and does not harm coding efficiency or increase complexity, and provides the opportunity fore reusing calculations in those implementations that might be able to do that.

• JVET-Q0365: Replace the LUT of GeoFilter[ ] by the proposed equation.

The proposed method is considered to be friendly for parallel processing in both hardware and software implementation, as a single equation can be used for per-sample weight calculation. (Basically no coding efficiency impact: 0.02% gain for RA, 0.06% loss for LB.)

Decision: Adopt per BoG recommendation.

BoG recommended for Track discussion:

• JVET-Q0422: Further discussion of aspect 2 and 3 in Track A or in an HLS session on high level control of coding tools.
  • See the notes for JVET-Q0482 regarding the signalling of number of merge candidates.

This was resolved as recorded for JVET-Q0798.

• Further discussion of candidate proposals on reducing the number of modes of geometric partitioning, as summarized in section 3.1.1 in the BoG report document. The claimed benefit is verification/validation efforts. The number of the combinations of GEO modes and CU shapes is 64*16 in the method previously agreed in the meeting. (The “common base” method had 82*19 combinations.) Decision: Stay with the 64-combination method from the CE, except disable GEO for 8x64 and 64x8.

Encoder will do 60 SATD check and 8 R-D checks.

• JVET-Q0437 (disabling MTS for geometric partition mode), only partial data for the encoder only restriction had been available in the BoG. It was noted that the transform and prediction are different stages of processing. No action was taken on this.

JVET-Q0761 BoG report of CE4 related contributions on inter prediction with geometric partitioning [H. Yang]

See section 6.4.