JVET-AP0052 AHG11: Improved H-DRF with Weighted Fusion and Optimized YUV Processing [X. Chen, J. Zhang, Z. Chen (Wuhan Univ.)]

This contribution builds on H-DRF (DRF at high operational point) from NNVC-16.0 by introducing a HardSigmoid-based weighted fusion layer, performing optical-flow computation directly in the YUV domain, optimizing quantization for model conversion, and replacing bilinear down/up sampling with Lanczos for YUV420/YUV444 conversion.

The coding gains of the proposed method are reported as follows:

1) Int16, SADL implementation, RA, compared to NNVC-15.0 anchor (NNIntra+LOP6):

Overall: -2.76% (Y), -2.98% (U), -3.09% (V) EncT: 142% DecT: 2420%

2) Int16, SADL implementation, RA, compared to NNVC-15.0 (H-DRF from NNVC-16.0 + NNIntra + LOP6):

Overall: -0.28% (Y), -1.79% (U), -3.33% (V) EncT: 100% DecT: 101%

It was reported that complexity is practically identical to current H-DRF in NNVC SW, but gives gain (see 2) above).

It was asked if there would be a chanc to avoid the conversion 4:2:0/4:4:4 and back? This might be beneficial in terms of complexity, and improve performance.

Investigate in EE. A similar approach should also be investigated for the lower-complexity DRF versions.