JVET-Q0447 Encoder estimation of weighted_prediction parameters [P. Bordes, T. Poirier, F. Le Léannec (InterDigital)]
The current VTM-7.0 encoder software implements 5 methods for estimating WP parameters.
The methods 0,1 use AC and DC of the original and reference pictures to derive WP parameters. The methods 2,3,4 use histograms of the original and reference pictures to refine the WP parameters computed with AC, DC.
All these methods use sum of absolute difference (SAD) between the original and the corrected reference samples to finally select the best WP parameters.
To be more consistent, it is proposed one alternative histogram-based method replacing SAD with histogram distortion for selecting the best WP parameters.
It is reported that compared to the VTM-7.0 method-2, the proposed method provides BD-rate gains of x.xx%(Y), x.xx%(U), x.xx%(V) / x.xx%(Y), x.xx%(U), x.xx%(V) in RA / LDB configurations with fade-to-black sequences, and x.xx%(Y), x.xx%(U), x.xx%(V) / x.xx%(Y), x.xx%(U), x.xx%(V) in RA / LDB configurations with fade-to-white sequences.
The encoder and decoder complexity are unchanged compared to the anchors.
Results are only reported with a modified method 2, where it is shown that if the modified method is applied for luma and chroma, a loss occurs in chroma. Therefore, the proponents suggest (in their variant B), to apply the histogram based method only for luma, and keep chroma unchanged. With that, the gain of modified method 2 versus VTM method 2, which gives 1.1/1.4% luma gain for RA, 0.8/0.6% in LD (cases fade-to-white/fade-to-black). There is also some chroma gain, probably because the overall rate becomes lower.
Decision (SW): Adopt JVET-Q0447 variant B (non CTC).