JVET-H0010 JVET AHG report: Denoising and adaptive quantization [R. Sjöberg, E. Alshina, S. Ikonin, A. Norkin, T. Wiegand]

The document summarizes the activities of the AHG on Denoising and adaptive quantization between the 7th and the 8th JVET meetings.

Two related contributions were noted

• H0047 Perceptually optimized QP adaptation and associated distortion measure:

This document describes an adaptive quantization algorithm and a corresponding distortion metric.

• H0055 Analysis of the HDR Objective Metrics Utility in Draft CfP on FVC:

This contribution reviews the HDR objective metrics listed in the draft CfP and compares the correlation between subjective and objective results done previously.

An investigation of denoising filters was done by Ericsson and results were shared on the e-mail reflector.

Four encoding settings for AI, RA, LDB and LDP were run:

• HM 16.15 on anchor test set
• JEM 6.0 on anchor test set
• HM 16.15 on denoised test set
• JEM 6.0 on denoised test set

The anchor test set consist of the SDR CTC sequences. The denoised test set consist of the same sequences but after applying denoising. The Avisynth tool MC_Spudsmod was used with the following settings: (frames=4, strength=1, sharpp=0, chroma=true). Further information can be found in the JVET-G0073 contribution from Torino.

The HM encodings were done using the HM-16.6-JEM-6.0\cfg\hm-16.13 configurations. Those give identical results as the HM anchor reported in JEM4-vs-JEM5-gcc62.xlsm. The BD-rate savings of JEM over HM were calculated. When comparing encoding 3 and 4 the denoised source was used for computing PSNR and MS-SSIM values.

The results can be found in tables in the AHG report. Full results can be found at http://phenix.int-evry.fr/jvet/doc_end_user/documents/7_Torino/comments/comment-2-JEM-vs-HEVC-on-denoised-sequences.zip. The differences in JEM-over-HM BD-rate savings between coding original sequences and coding denoised sequences were reported to be fairly small. The results are tabulated in the AHG report.

In the review of the AHG report, it was noted that the processing done for these experiments involved an 8-bit “bottleneck”, so that a readily available software package could be used as-is.

A suggestion to include anchors outside of competition to the CfP subjective tests for SDR-HD1 or SDR-UHD1 using adaptive quantization and denoising was made on the reflector. The motivation for that was that it would quantify the MOS improvements from using adaptive QP, denoising and renoising. Also, it would provide MOS scores from encoders that are rather different compared to the anchors which can be used for a later evaluation of different quality metrics.

No conclusion on the suggestion was reached by the AHG.

In the draft CfP, denoising is prohibited in all cases. QP adaptation related to average local brightness is allowed for HDR content (and is used with a specific brightness adaptation rule in the anchors). QP adaptation as a function of geometric position (not customized to the content) is allowed for the 360° cases.

Analysis, development and improvement of JEM (1)

Contributions in this category were discussed Sunday 22 Oct. 1100–1130 (chaired by JRO & GJS).