JVET-AJ0010 JVET AHG report: Encoding algorithm optimization (AHG10) [P. de Lagrange, A. Duenas, R. Sjöberg, A. Tourapis (AHG chairs)]

• Activity

The regular JVET e-mail reflector was used for discussions (jvet@lists.rwth-aachen.de).

Teleconferences were held on August 20^th, September 5^th, September 26^th and October 14^th, and a few related organizational and reporting e-mails were sent to the JVET reflector. No other e-mail related to AHG10 activity was sent during the AHG period.

• Related contributions

Four contributions, not including cross-checks, were identified relating to AHG10, and summarized in the following sections.

• Scalability
• JVET-AJ0042 - AHG10: Teleconference on encoder optimization for multi-layer coding

This document summarizes discussions during teleconferences on encoder optimization for multi-layer coding, focusing on the mandate related to multi-layer coding.

The work is related to an issue found during visual tests of spatial scalability in Geneva (31^st meeting), where the VTM encoder could create an enhancement layer that degrades visual quality (see JVET-AE0288), introducing coding artefacts that outweigh the benefits of detail enhancement. The situation was improved at the following meeting (see JVET-AF0186 and JVET-AF0311) but is still not satisfying.

The goal of the teleconference was to organize encodings and preliminary checks (both objective metrics and informal visual reports) with different configuration and/or new ideas and spread the load, so that meaningful visual tests can be conducted with AHG4/AG5.

Suggested tests included evaluating the impact of adaptive QP, LMCS, layer rate allocation, base layer MCTF, DMVR, quantization matrices, enhancement layer GOP structure parameters like deblocking strength, intra QP offset, and reference picture lists. It was also suggested to study the pumping effect, and improve the RDO around inter-layer reference selection.

Some planned improvements of software tools, including bug fixes, were made and merged into the VTM, e.g. configuration of inter-layer reference pictures in the GOP structure, upscale/downscale. A bug was fixed around LMCS-related data accidentally shared across layers. It is noted that the use of --DebugBitstream greatly speeds up multi-layer encodings reusing the same base layer.

Significant progress was made on the issue under study, using both configuration and RDO changes for the enhancement layer. The changes are described in JVET-AJ0067 and JVET-AJ0225.

• JVET-AJ0067 - AHG10: Suggested updates of settings for scalable VVC coding

This contribution suggests three modifications to the settings for the enhancement layer for scalable VVC coding:

• omit the base layer reference picture from L1 in the explicit reference picture configurations such that its only present in L0
• decrease the IntraQPOffset to -4 from -3 to put some more emphasis on details
• use unbiased deblocking settings for the enhancement layer such that deblocking can be used more often.

Using the test set and coding conditions of JVET-AI2021 and JVET-AJ0042, the combined BD-rate impact is reported to be -0,78%/-0,59%/-0,55% for Y/U/V PSNR. It is asserted that visual quality also is slightly improved.

• JVET-AJ0225 - AHG3/AHG10: RDO adjustments for inter-layer predictions

This contribution proposes adjustments of the encoder selection of coding modes for dual-layer coding in order to prevent quality degradation of the enhancement layer compared to the base layer. The rate-distortion (RD) criterion is adjusted to favor the selection of inter-layer coding modes over intra-layer coding modes using a modified RD cost. Modifications are also proposed to make sure that inter-layer coding modes are included in the rate distortion optimization process.

It is asserted that visual quality is improved in areas where local degradations compared to the base layer were observed.

Using the test set and coding conditions of JVET-AI2021 and JVET-AJ0042, the BD-rate is reported to be +0.48%/+0.51%/+0.54% for Y/U/V PSNR. The BD-rate of a further combined test with JVET-AI0067 are reported to be -0.37%/-0.24%/-0.01% for Y/U/V PSNR.

• VTM and ECM speed-up related to MTT split decisions
• JVET-AJ0226 - AHG12/AHG10: Using CABAC cost for MTT split modes selection

This contribution presents a modification of the MTT split modes early termination method in ECM/VTM. The cost of coding the CU without splitting the CU at all is compared to a threshold based on the QP value and the CABAC cost for coding a certain MTT split mode flag. If the cost is higher than the threshold, the corresponding MTT split mode will not be considered.

Encoding time reduction is reported (-4% for ECM, -6% for VTM, with a stronger impact for QP22: -13% and -16%), with small BD-rate losses: +0.03/0.08% for ECM RA/LDB, +0.06/0.12% for VTM.

• Recommendation

The AHG recommended that the related input contributions be reviewed, that expert viewings be conducted at the current meeting to assess the improvement brought by JVET-AJ0067 and JVET-AJ0225 to dual-layer coding, and to further continue the study of encoding algorithm optimizations in JVET.