JVET-AP0010 JVET AHG report: Encoding algorithm optimization (AHG10) [K. Andersson, P. de Lagrange, A. Duenas (co-chairs), T. Ikai, T. Solovyev, A. Tourapis (vice chairs)]
- Related contributions
Following contributions were identified relating to AHG10 and summarized in the following sections.
JVET-AP0070: AHG10: Additional random-access configuration with reduced delay
This document suggests adding an additional random-access configuration file based on a length 8 hierarchical GOP structure. The document also suggests updating the reference software with some additional parameters to control the QP every N:th frame to mimic quantization for a larger hierarchical GOP structure. The overall impact on BDR-rate compared to CTC using a length 32 hierarchical GOP structure is reported to be +7.88%/+13.08%/13.35% (Y/U/V) with encoding time of 99.7% and decoding time of 100.3%. Additionally, results for also enabling GOPbasedRPR and BIM=2 reports an overall BDR impact of +6.02%/+10.84%/+10.92% with encoding time 98.1% and decoding time 92.9%. The additional results indicates that both GOPbasedRPR and BIM also works together with the hierarchical GOP of length 8.
JVET-AP0191: [AHG16] VTM software extension for counting and constraining per-CU RDOs
This document recommends creating a AHG16 git repository for development of the complexity metrics as well as the hardware encoding mimicking simulation platform, e.g., with per-CU RDO number counting and constraints. The aim is to make the next-generation video coding more hardware friendly for both encoding and decoding.
In comparison to CTC (VTM-23.9) for random-access configuration the overall BDR impact of a limitation of the number of RDO per CU for intra/inter is reported to be:
- 10/10: 1.93%/1.20%/1.66% (Y/U/V) with encoding time of 88% and decoding time of 100%.
- 5/5: 4.80%/6.88%/7.31% (Y/U/V) with encoding time of 78% and decoding time of 100%.
JVET-AP0192: [AHG19] Evaluating VTM under hardware constraints
This document recommends establishing a test configuration that considers hardware encoding constraints, for the development of next-generation standard. The GOP8 random-access configuration file proposed in the contribution is recommended to be a starting point.
The overall BDR impact of all tested encoding constraints in the document compared to CTC is 24.38%/35.22%/36.55% (Y/U/V) with encoding time of 15% and decoding time of 99% compared to VTM CTC (VTM-23.9). The test is using a GOP 8 random-access configuration with a limitation to 2 reference pictures, cfg/alternative-addon/reduced_runtime3.cfg, and a limitation to the number of intra and inter RDO’s to 10 respectively.
In comparison to an anchor using cfg/alternative-addon/reduced_runtime3.cfg the overall BDR impact of the GOP 8 random-access configuration with a limitation to 2 reference pictures is 12.12%/19.97%/19.69% (Y/U/V) with encoding time of 98% and decoding time of 101%.
JVET-AP0237: AHG17: Improved VTM configuration for 0.2x runtime target
This document reports results for a modified configuration for 0.2x runtime target with BDR rate impact of 3-4% on top of VTM-24.0 for the CfE test set and it also asserts visual improvements compared to the current 0.2x configuration. The document also provides rate matched settings for the VTM default anchor, the VTM RPR comparison point with BIM=2, the old 0.2x runtime target and the proposed 0.2x runtime target. The proposed modifications for the 0.2x runtime target apply for both random-access and low-delay configuration and are applied on top of the corresponding default configurations (Note: LMCSEnable settings are overridden by HDR specific settings):
Random-access:
MergeRdCandQuotaRegularSmallBlk : 3, MergeRdCandQuotaRegular : 3, MaxNumGeoCand : 4,
UseNonLinearAlfLuma : 0, AffineAmvp : 0, MaxNumAffineMergeCand : 4, SplitPredictAdaptMode : 2,
AdaptBypassAffineMe : 1, MergeRdCandQuotaGpm : 3, TransformSkip : 0, MaxMTTHierarchyDepth : 1, MaxMTTHierarchyDepthISliceL : 2, MaxMTTHierarchyDepthISliceC : 2, MTS : 4, SBT : 0, LFNST : 1, MMVD : 0, MaxNumMergeCand : 6, BCW : 0, Geo : 1, AllowDisFracMMVD : 0, AffineAmvr : 0, LMCSEnable : 0, MIP : 0, SMVD : 0, ChromaTS : 0, MaxMergeRdCandNumTotal : 4, AffineAmvrEncOpt : 0
Low-delay:
AffineAmvp : 0, UseNonLinearAlfLuma : 0, MergeRdCandQuotaGpm : 5, SplitPredictAdaptMode : 2,
AdaptBypassAffineMe : 1, TransformSkip : 0, TemporalFilter : 0, MaxMTTHierarchyDepth : 1, SBT : 0,
BCW : 0, AllowDisFracMMVD : 0, ChromaTS : 0, MaxMergeRdCandNumTotal : 6
Recommendations
The AHG recommended that the related input contributions are reviewed, and to further continue the study of encoding algorithm optimizations in JVET.