JVET-AC0010 JVET AHG report: Encoding algorithm optimization (AHG10) [P. de Lagrange, A. Duenas, R. Sjöberg, A. Tourapis (AHG chairs)]

1. Related contributions

A total of 5 contributions, not including cross-checks, were identified relating to AHG10, and summarized in the following sections.

It is noted that some of the mandates of AHG15 are related to test conditions and encoder optimization, and that some level of coordination would probably be beneficial.

Test conditions

JVET-AC0078 – EE1-2.1: Updates on RPR encoder and post-filter

The JVET-AB0102 contribution listed in the AHG10 report for previous JVET meeting proposed using a secondary anchor for EE1 super-resolution tests, that includes more intermediate resolutions (adding 4/5 ratio), GOP-based decision of coding resolution based on QP and high-frequency content in video source, and longer-tap filters for upscale.

This contribution reports test results in the EE1 context for these three aspects. GOP-based decision brings 0.7% luma gain in RA using only 1/1 and 1/2 resolutions, 1.23% when adding 2/3 and 4/5 resolutions, and longer tap filters bring around 0.3% chroma gain.

JVET-AC0096 – AHG10/12: Suggestion for new CTC for RPR in VTM and ECM

This contribution proposes four changes to RPR CTCs for both VTM and ECM, to exercise the tool more and adjust viewing conditions:

• Introduce more resolution changes: more often, and more resolution variants (switching every 32 frames and alternating between full resolution and one of the reduced resolutions: 4/5, 2/3, 1/2)
• Harmonize bitrates between the different resolutions by adjusting QPs depending on resolution (-6 for 1/2, -4 for 2/3, -2 for 4/5)
• Evaluate performance by using PSNR of upscaled reconstructed pictures only
• Test in RA configuration in addition to LDB

Performance numbers are reported using the proposed test conditions, for both VTM and ECM, and also for GOP-based RPR using VTM.

JVET-AC0138 – Adjusting luma/chroma BD-rate balance in ECM

This contribution proposes updates of ECM CTCs to balance luma/chroma gains compared to VTM. More specifically, adjustments to “CbQpOffset”, “CrQpOffset” and “LMCSOffset” parameters are proposed, depending on SDR/HDR and AI/others.

For AI / RA / LDB where imbalance was around +9% / +6% / +10% gain for chroma, imbalance is reduced to +0% / +4% / +5% respectively, with luma gain improvement of +0.9% / +0.2% / +0.16%.

JVET-AC0149 – AHG10/12: Reduced I-frame QP for RA

This contribution proposes to change the QP offset for intra frames in RA CTCs for both VTM and ECM: changing the “IntraQPOffset” config parameter from -3 to -4 is reported to bring 0.28%/2.62%/2.76% Y/Cb/Cr BDRate gains in the VTM, and similar gains for ECM. Subjective benefits are reported for static content, with no subjective problems for other content.

Encoder optimizations

JVET-AC0139 – AHG10: Encoder Optimization of VTM Merge Functions

In this contribution, an encoder only optimization on merge is proposed for VTM software. 97%, 99% and 92% encoding time is reported with 0.02%, 0.01% and 0.03% luma BD-rate loss for RA, LDB and LDP configurations, respectively.

Three merge functions on regular merge, sub-block merge and geometric merge are replaced with one unified merge function, which is said to remove redundant code. The code changes also include a new data structure to hold candidate information and prediction buffers, and memory optimization of prediction buffers. The maximum number of RDO checks after SATD-based candidate pruning is capped by a new configuration parameter (Num_MaxRdCfg). The “Best Skip based Fast Merge” method is removed based on infrequent use in CTCs, marginal impact on encoding time, and losses in specific conditions.

Recommendation

The AHG recommended that the related input contributions be reviewed and to further continue the study of encoding algorithm optimizations in JVET.