JVET-AD0196 EE2-related: On Iterative BDOF and EE2-2.6 Improvement [M. Salehifar, Y. He, K. Zhang, H. Liu, L. Zhang (Bytedance)]

In test EE2-2.6 (JVET-AD0195), high-precision MV refinement for BDOF is introduced, which includes accurate BDOF parameter calculation, added weights, as well as adaptive BDOF DMVR subblock size.

This proposal introduces some improvement on top of EE2-2.6, including adding iteration for BDOF DMVR for qualified blocks.

On top of the ECM-8.0 and EE2-2.6, simulation result of the proposed method is reported as:

Compared with ECM-8.0:

RA: -0.33%, -0.34%, -0.38%, 103%, 106%.

Compared with EE2-2.6:

RA: -0.13%, -0.15%, -0.20%, 103%, 105%.

Most of the gain is from iterative BDOF. One additional iteration was used.

Increasing number of iterations gave more gain for some sequences, but would also increase runtimes.

Proponents see options to reduce runtime (current tradeoff with compression benefit not good)

It was agreed to investigate this in an EE. Also to investigate the benefit of the different elements, and possibility of runtime reduction (both encoder and decoder).