JVET-V0056 [AHG10] GOP-based temporal filter improvements [P. Wennersten, C. Hollmann, J. Ström (Ericsson)]
This contribution proposes a set of changes to the GOP-based temporal filter in VTM that aim at improving the BD-rate performance. These changes were first proposed in JVET-U0056.
The average Y/U/V BD-rates for VTM-12 CTC are reported to be −1.4%/−1.0%/−1.1% for RA. All BDR numbers were computed using unfiltered source sequences as reference.
In addition, the same changes were implemented in HM where the BD-rates are -1.0%/-1.0%/-1.0%.
Visual inspection of the resulting sequences was reportedly carried out with no problems reported.
It is proposed to adopt the proposed changes into the VTM and HM software.
Version 2 of this document adds results for HDR.
Changes: Block error weighting (stronger filter when MC error larger), Block frequency weighting (stronger filter when higher frequencies present, Reference frame weighting.
Also a bug fix regarding MC at picture boundaries.
Subjective viewing: Equal or better quality was found by proponents.
Question: Would there be more headroom for subjective improvement? A: Likely by still stronger filtering.
Question: What happens in case of illumination changes? Might be appropriate applying a weight in the MC stage.
Decision (SW/CTC): Adopt (both for VTM and HM), and also use in the EEs.
The other items above for further study.
It was further suggested to investigate potential need for changing the filter weights in the EEs. A first step on this could be to investigate if the gain that MCTF gives for VTM is similarly preserved for the EE configuration.