Back to Search Document details
29th Meeting: by teleconference, DE, January 2023 2023-01-18 03:12
Non-EE2: SGPM without blending
Abstract
In ECM-7.0, spatial geometric partitioning mode (SGPM) uses a pre-defined blending width based on the block size, without encoder selection or signalling. The blending width could be 16 samples on each side of the partitioning line, no matter a block is of camera-captured content or screen-content. In this contribution, it is proposed to allow SGPM without blending for SCC.
JVET-AC0189 Non-EE2: SGPM without blending [Z. Deng, L. Zhang, K. Zhang, L. Zhao (Bytedance)]

In ECM-7.0, spatial geometric partitioning mode (SGPM) uses a pre-defined blending width based on the block size, without encoder selection or signalling. The blending width could be 16 samples on each side of the partitioning line, no matter a block is of camera-captured content or screen-content. In this contribution, it is proposed to allow SGPM without blending for SCC.

Simulation results reportedly show BD-rate reduction for Y, U, V components as follows.

AI configuration:

Class F: -0.25%, -0.14%, -0.18%; 100% enc time, 98% dec time

Class TGM: -0.38%, -0.31%, -0.35%; 99% enc time, 100% dec time

RA configuration:

Class F: -0.24%, -0.34%, -0.34%; 100% enc time, 100% dec time

Class TGM: -0.12%, -0.26%, -0.19%; 100% enc time, 100% dec time

An updated presentation deck was requested to be uploaded.

SGPM tool-off results in ECM7 for SCC classes:

Class F: 0.15%/0.05% for AI/RA

Class TGM: 0.15%/0.08% for AI/RA

For natural sequences, it is reported by the proponent that SGPM tool-off shows 0.24%/0.12% for AI/RA.

It was commented that some blending will still occur when the proposed flag is set to 1, if the SGPM block has partition angle that is not completely horizontal or vertical, but the blending area will be much narrower than what is currently used.

It was commented by several experts that excessive blending for SCC content could be harmful, and this proposal solves that problem. Several experts, including the original proponent of SGPM, supported adopting this change for SCC content.

It was commented that the proposal uses PPS flag rather than SPS flag because it could allow a finer granularity control of the blending process (e.g. in the case when some natural pictures are followed by screen content pictures).

The proponent further suggested to set this flag to 1 in the case of SCC CTC.

It was suggested that further encoder optimization could use this proposed method to bring gains to some natural content or mixed content by switching the flag picture by picture.

Decision: Adopt JVET-AC0189.

Decision (CTC): set the PPS flag to 1 for class F and class TGM.

Decisions
adopted
Adopt JVET-AC0189
Citation