JVET-AD0010 JVET AHG report: Encoding algorithm optimization (AHG10) [P. de Lagrange, A. Duenas, R. Sjöberg, A. Tourapis (AHG chairs)]
Related contributions
A total of 7 contributions, not including cross-checks, are identified relating to AHG10, and summarized in the following sections.
It is noted that some of the mandates of AHG8 are related to test conditions and encoder optimization, so some of the contribution relating to AHG8 can also be of interest for AHG10.
JVET-AD0045 - AHG10: Encoder MV selections and DMVR revisited
This contribution describes an encoder method to decrease visual artefacts related to DMVR on sub-block boundaries, with lower BDRate impact than what was proposed in JVET-W0061. It consists in avoiding usage of DMVR when motion derivation is more likely to be unreliable and problems could propagate, e.g. in low temporal layers when reference pictures are far, and/or frame rate is low, and when artefacts are likely (spatial activity at boundaries). There are also conditions related to block size and QP. Compared to JVET-W0061, BDRate losses are reduced from 0.2% to around 0.04% (SDR CTCs) / 0.08% (HDR CTCs). In the context of GOP-based RPR and higher QPs, losses are around 0.5% compared to 2.3% for DMVR-off.
Bug fix
JVET-AD0129 - AHG10: Improvement of Input Video Padding in VTM
This contribution describes a bug fix in the VTM when the right side of the image is padded to reach a multiple of 8 or minimum CU size, before entering the encoder. This condition is not triggered by the CTCs. When tested with relevant picture sizes (cropped from the CTC sequences), impact on BDRate is significant (around -0.35% for RA).
Lambda adjustment (RDO)
JVET-AD0133 - AHG10: Lambda-QP Relationship Fix for Slice-level Multi-QP Optimization
This contribution proposes fixes to the frame lambda computation in case of slice-level QP RDO. The lambda is derived from m_vdRdPicLambda instead of a (potentially inconsistent) formula, and is made independent from the QP sweep. Also a conflict with CU results cache is solved by disabling the cache in that case. BDRate impact as high as -10% is reported in random access.
JVET-AD0136 - AhG10: Lagrange multiplier optimization for chroma ALF and CCALF
This contribution proposes to adjust the lambda for the RDO of chroma ALF when luma ALF coefficients are to be transmitted, to account for a reduced overhead of the combined signalling, and also encourage combined signalling (resulting in more chroma filter diversity). Chroma gains are reported.
Low-delay, GDR
JVET-AD0206 - AHG 12: CABAC Initialization for GDR Pictures
This contribution reports that inter-picture CABAC state inheritance as implemented in ECM8.0 is currently incompatible with GDR as it can cause mismatch at recovery point. It is proposed to enable that feature by always forcing CABAC initialization for GDR pictures. BDRate difference of around -0.5% is reported over TempCabacInit=0 (in low delay configuration with GDR enabled).
Local QP adaptation
JVET-AD0138 - [AHG8] QPA with low activity threshold for machine task
This contribution proposes a method to prevent the local QP adaptation to lower the QP too much on smooth area, by applying a low-activity threshold. This type of technique is generally useful, even outside of machine analysis tasks, for example to avoid catching noise in intra picture, causing intra pumping or “dirty window” effect.
Adaptive resolution
JVET-AD0169 - AHG12: RPR filters for scale factors below 1.5x
This contribution proposes to:
- Change non-normative downscaling filters for scaling ratios below 1.33x to larger bandwidth filters, providing sharper downscaled input picture (beneficial to upscale RPR)
- Replace the normative downscale RPR filters in a similar way (beneficial to downscale RPR)
Reported combined BD-Rate difference with ECM-8.0 is around -9% (for 1.25x) and -13% (for 1.33x), using RPR random-access test conditions, whereas non-normative changes alone result in around -8% (for 1.25x) and -10.5% (for 1.33x), which seem to indicate that sharp downscaled pictures are of primary importance for adaptive resolution performance.
The proposed downscaling filters could be used in the general case, even for the VTM and scalability cases.
Recommendations
The AHG recommended that the related input contributions be reviewed and to further continue the study of encoding algorithm optimizations in JVET.