Back to Search Document details
15th Meeting: Gothenburg, SE, July 2019 2019-06-26 09:46
CE4-related: Simplifications on PROF
Authors: Huanbang Chen Huawei Hyeong-Gyun Yang Huawei
Abstract
In CE4-2.1, Prediction refinement with optical flow (PROF) for affine mode was tested. It’s reported the luma BD-rate for RA and LD respectively are -0.48%, and -0.32% in test a. In this contribution, two simplified aspects for PROF are proposed. The first aspect is to apply PROF to affine coded block with uni-prediction for reducing the complexity of PROF. In the second aspect, the refinement offset is clipped within [-4092, 4091] to ensure the refined value is within 16-bit signed range.
Meeting Notes
JVET-O0664 CE4-related: Simplifications on PROF [H. Chen, H. Yang (Huawei)]

JVET-O0664 CE4-related: Simplifications on PROF [H. Chen, H. Yang (Huawei)]

In the second aspect, the refinement offset is clipped within [-4096, 4095] to ensure the refined value is within 16-bit signed range.

Simulation results for aspect 2 reportedly show that the luma BD-rate -0.48% and -0.30% in RA and LB cases, respectively, when compared to VTM-5.0.

In current design, BDOF has no clipping since the offset is applied in the final prediction signal stage. In CE4-2.1a that PROF has overflow issue since it is applied in the prediction stage of each reference picture list.

If the uniform design of BDOF and PROF is adopted in this meeting cycle, there is no need to take action on this contribution. Otherwise, the bug-fix proposed in this contribution is recommended for adoption.

In the follow-up discussion in track B, if was agreed that the clipping is not needed, since the prediction before adding the offset is not using the full 16 bit range. No action.

Cited:
JVET-O0732 X
ARCHIVE :
JVET-O0664-v1.docx JVET-O0664-WD.docx JVET-O0664_test1.XLSM JVET-O0664_test2.XLSM
Decisions
JVET-O0664 CE4-related: Simplifications on PROF [H. Chen, H. Yang (Huawei)]
In the follow-up discussion in track B, if was agreed that the clipping is not needed, since the prediction before adding the offset is not using the full 16 bit range. No action.
Citation