JVET-U0010 JVET AHG report: Encoding algorithm optimization (AHG10) [A. Duenas, A. Tourapis, A. Norkin, R. Sjöberg]

At the 20th JVET meeting, the AHG on Encoding algorithm optimizations was established with the following mandates:

• Study the impact of using techniques such as GOP structures, GDR, LMCS and perceptually optimized adaptive quantization for encoder optimization.
• Study encoding techniques of optimization for objective quality metrics and their relationship to subjective quality.
• Particularly consider neural network-based encoding optimization technologies.
• Study the impact of adaptive quantization.
• Investigate other methods of improving objective and/or subjective quality, including adaptive coding structures and multi-pass encoding.
• Study methods of rate control and rate-distortion optimization and their impact on performance, subjective and objective quality.

The regular JVET e-mail reflector was used for discussions (jvet@lists.rwth-aachen.de). No e-mail related to AHG10 activity was sent to the JVET reflector during the AHG period.

The following input documents were identified to be related to this AHG:

• JVET-U0056: [AHG10] GOP-based temporal filter improvements

This contribution proposes a set of changes to the GOP-based temporal filter that is included in the VTM that aim at improving PSNR BD-rate performance.

The proposed changes were reportedly tested under the VTM-11.0 RA common test conditions (CTC) The filter is not used for AI or LD test conditions. On average a Y/U/V BD-rate performance improvement of −1.3%/−1.0%/−1.1% for the RA CTC compared to the current filtering method and an average BD-rate improvement of -5.4%, -8.2%, and 8.0% compared to no use of filtering are reported. More detailed results are shown in the next section.

The proposal was implemented on top of VTM-11.0. The method is currently applied for the RA coding conditions only. It would be of interest if the scheme could be extended to other configurations or if finer control of the filtering parameters was included.

BD-rate performance relative to the current scheme in VTM-11.0 (MCTF = 1) was reported as follows

In addition, the contribution includes the following results when compared to VTM-11.0 without temporal filtering:

It was noted that the proposed method shows a complexity reduction, in addition to the coding benefits such a scheme is introducing, compared to the use of no filtering. Furthermore, it is indicated also in the text that there seems to be no discernible impact on encoding complexity. It would be of interest to analyse such behaviour and see if such relates to the improved temporal characteristics, and thus implications in motion estimation and mode decision, of the coded sequences. Additional analysis at different operation/bitrate points may also be advisable.

• JVET-U0081: [AHG10] ALF filter optimization with filter strength target

The VTM encoder currently has the possibility to control the amount of ALF filtering by changing the filter strength parameters. The current VTM behaviour is that the optimal filters are calculated and are then scaled by the filter strength parameters (decreasing the magnitude) as part of the quantization. That approach can, together with disabling pre-defined filters and omitting the refinement step of the quantized coefficients, reduce the filter strength for ALF according to the filter strength parameters. This proposal suggests including the filter strength target parameter in the ALF filter optimization. This approach favours filters that have a reduced strength according to the filter strength target in all stages of the VTM’s filter optimization, i.e., during merging, refinement of quantized coefficients and when using pre-defined filter coefficients, but does not guarantee a reduction of filter strength according to the target parameter. Separate parameters for luma and chroma ALF are also suggested.

A BD-PSNR rate impact with filter strength target parameters of 0.875 for both luma and chroma is reported to be equal to 0.04%/-0.03%/-0.06% for the AI/RA/LDB test conditions respectively. A BD-PSNR rate impact with filter strength target parameter of 0.875 only for luma of 0.04%/-0.02%/-0.09% for the AI/RA/LDB test conditions respectively is also reported. The authors also provide results for the current approach in the VTM for the case of ALFStrength equal to 0.875. In this case the BD-PSNR rate impact is equal to 0.10% for AI and 0.07% for RA test conditions. LDB results were not included in the latest version.

It was claimed that the proposed approach can favour filters according to the filter strength target and that in some cases it can improve the subjective quality of VTM for inter predictive coding.

• JVET-U0097: GDR Software

This contribution includes a software update for includes gradual decoder refresh (GDR) support in VTM11.0. The authors claim that since the VVC specification has now been finalized, it was also possible to make the corresponding GDR software, based upon VTM11.0, also more stabilized. Therefore, they propose integrating this implementation into the VTM source code package.

The authors mention that the following features, relating to GDR, are supported by this implementation:

• • Encoding input video sequences into GDR bitstreams with the following features:
    • Flexible GDR period configured through parameter set by user,
    • The first picture can be either IDR or GDR picture (configurable),
    • SPS/PPS/APS are signalled at each GDR picture,
    • Similar number of bits per picture, implying the delay is as low as one frame interval,
    • Progressive intra refresh over a GDR period with even distribution of the forced intra areas over pictures within the GDR period using virtual boundary syntax in picture header,
    • Necessary (encoding) constraints on coding tools to prevent the leaks,
    • Exact match at the recovery point (or leak-free).
    • MD5 can be turned on or off for GDR pictures and recovering pictures (configurable),
    • ph_pic_output_flag may be signalled for GDR pictures and recovering pictures.
  • Decoding the leak-free GDR bitstream using the VTM decoder.
    • If MD5 is off for GDR pictures and recovering pictures, decoder will ignore hash check for those pictures and only check the hashes for non GDR/recovering pictures.

There was no mention of whether a similar implementation could also be provided for previous standards and their reference implementations (e.g. the HM reference software of HEVC). Such implementations may, however, be desirable.

It is noted that this contribution is only for encoder improvement, not related to the GDR conformance issue discussed under AHG5.

• JVET-U0103: AHG8: SIMD support for VTM software at high bit-depth coding

This contribution provides software a patch that enables single instruction multiple data (SIMD) support for the VTM-11.0 software when RExt__HIGH_BIT_DEPTH_SUPPORT is enabled for high bit-depth. Compared to the VTM-11.0 12/16-bit anchor, simulation results reportedly show that both the encoding and decoding times are reduced by about 40% for RA and LD configurations, with bit-exact BD-rate results.

The AHG recommended that the related input contributions be reviewed and recommended to further continue the study of encoding algorithm optimizations in JVET. In addition, it was recommended to coordinate with the appropriate CE group for high-bit-depth coding, the review of contribution JVET-U0070 on encoder optimization for Rice parameter derivation.