JVET-AG0010 JVET AHG report: Encoding algorithm optimization (AHG10) [P. de Lagrange, A. Duenas, R. Sjöberg, A. Tourapis (AHG chairs)]
Related contributions
Three contributions in total, not including cross-checks, are identified relating to AHG10, and summarized in the following sections.
JVET-AG0116 – AHG12: GOP-based RPR encoder control for ECM
This contribution is a follow-up of JVET-AF0058 that was porting the adaptive resolution algorithm available in the VTM (using the “GOPBasedRPR” configuration parameter) to ECM. Half the gain compared to the VTM case was reported, however such gain still considered as significant (around 1.5%; more than 3% in class A). Benefit was mostly reported for high resolution content and high QPs. The proponents were asked to provide full CTC results: it turns out that the objective performance results are almost neutral for the CTC (gains are still observed in class A), but the proponent reports some visual benefit.
Local QP optimization
JVET-AG0055 – CTU-Level Lagrange Multiplier and QP Adaptation for VVC Low-Delay Configuration
This contribution is a follow-up of JVET-AF0089, that proposes a method that tries to give more quality weight to local areas that are likely to be reused in other picture (through temporal prediction). It does so by estimating a “distortion propagation factor” with a fast 1st inter pass.
During the 32nd JVET meeting, several experts commented that the contribution was interesting, and further optimization could be applied to reduce loss, optimize the interaction with BIM (Block Importance Mapping, a single-pass MCTF-dependent QP adaptation in the same spirit, already available in the VTM and the HM), etc. It was suggested for the proponents to also examine the impact on subjective quality.
In this follow-up version, scene cut detection has been introduced to mitigate the effect on the computation of the “distortion propagation factor”. BD-rate difference of -4% is reported in low-delay configurations, with a moderate increase of encoding time (around 2 to 3%).
A comparison with BIM is given, notably stating that BIM is active only on 1 picture out of 8 (the rate at which MCTF is active), while the method in this contribution impacts every picture, and has finer control over the Lagrange multiplier. As a reminder, the gains of BIM are around 2.3% in the low-delay and 2% in the random-access test conditions.
Residual spatial weighting
JVET-AG0217 – Reduced residual encoding in VVC for machine consumption
This contribution reports about a residual spatial weighting method (attenuating the residuals by 75% in the center of a coding block -when larger than 4 in either dimension- but preserving DC), which is said to provide gains for machine related tasks (a few % BD-rate (mAP); incomplete results). BD-rate (PSNR) loss is around 0.45%.
Recommendation
The AHG recommended that the related input contributions be reviewed and to further continue the study of encoding algorithm optimizations in JVET.