JVET-AC0138 Adjusting luma/chroma BD-rate balance in ECM [E. François, Y. Chen, C. Salmon-Legagneur (InterDigital)]

This contribution reports experiments based on ECM-7.0 aiming at obtaining a better repartition of BD-rate variations between luma and chroma, when using VTM-11.0ecm6 as reference. As a result, the contribution tests adjustments of the configuration parameters “CbQpOffset”, “CrQpOffset” and “LMCSOffset” for the configurations of the ECM CTCs for SDR and HDR contents.

Several encoder configurations are tested for SDR contents, leading to the following AI/RA/LB BD-rate PSNR results over VTM-11.0ecm6:

• ECM-7.0: AI: -9.1%/ -18.0% / -19.0%; RA: -19.2%/ -25.4%/ -26.7%; LB: -16.2%/ -26.9%/ -26.3%
• config. 1: AI: -10.1%/ -10.2% / -11.7%; RA: -19.4%/ -22.9%/ -24.3%; LB: -16.4%/ -22.3%/ -22.1%
• config. 2: AI: n/a ; RA: -20.4%/ -15.4%/ -17.5%; LB: -16.7%/ -16.1%/ -14.7%

One encoder configuration is tested for HDR contents, leading to the following AI/RA results over VTM-11.0ecm6 (provided numbers are BD-rate wPSNR for H1, BD-rate PSNR for H2):

• ECM-7.0: AI/H1: -7.37% / -25.17% / -30.93%; AI/H2: -7.23% / -17.05% / -20.56%;
• config: AI/H1: -8.65% / -8.52% / -15.25%; AI/H2: -7.93% / -7.75% / -11.82%;
• ECM-7.0: RA/H1: -14.35% / -33.89% / -39.81%; RA/H2: -13.31%/ -26.59%/ -30.97%*
• config: RA/H1: -15.39% / -18.47% / -24.97%; RA/H2: -13.89%/ -14.79%/ -20.05%*

* H2 (HLG) RA results were not yet fully available – one missing rate point (SunsetBeach, QP 22) was replaced by H2 RA results previously generated, but with wrong config files (per_class classA_randomaccess.cfg omitted)

It was commented that a similar contribution (JVET-Y0223) had been brought previously (which included an additional knot over “QPOutValCb” to control the luma/chroma balance), and concerns were expressed at the time regarding the potential impact to subjective quality. It was mentioned by the proponents that subjective quality had been checked on a few sequences, and no obvious artefacts were found.

It was discussed that the current imbalance could be due to the adoption of more chroma tools (e.g. cross component prediction tools) in ECM.

It was commented that in some configurations/content, there seemed to be a rather large drop in chroma gains and in return a relatively small increase in luma gains.

It was commented that this contribution was intended as for information only, in order to show how more-balanced gains between luma and chroma in the current ECM could look like, and to show if 20+% luma gain is to be achieved, how much chroma gains could still remain.

Further study was encouraged, including other methods to achieve better balance between luma and chroma (such as using QPCMappingTable).