JVET-D0192 BoG Report on Extended Colour Volume Material [A. Segall]

The BoG met on October 18, 2016 to address the following mandates:

• Review EE9/AH7 related contributions
• Discuss and agree on conversion methods
• Discuss and agree on test procedures for extended colour volume materials.

Conversion methods:

Overview: The goal of the conversion method discussion is to identify the target container or containers for studying the coding performance of extended colour volume material, and how content in other containers would be converted to the identified target containers.

It was suggested to consider the BT.2020 colour primaries in the target container

It was suggested that PQ should be considered as a transfer function.

It was suggested that HLG could be considered as a transfer function. However, this could be done as a second step and later activity. One participant suggested that we may be able to convert from PQ to HLG.

Recommendation of BoG: Use a container with BT.2020 colour primaries and the PQ transfer function.

When converting legacy (SDR) content to the target container, one suggestion was to map the RGB 8-bit value of 235 to 100 cd/m2 and map the RGB 8-bit value of 16 to 0.01 cd/m2, as defined in BT.2035.

Note that it will be necessary to confirm the scaling. It was commented that perhaps as a way to accelerate the study, it would be useful to generate multiple representations with different cd/m2 value targets. For example, 300 and/or 400 cd/m².

There was clarification discussion about the meaning of scaling. It was suggested that this would be a multiplication. Alternatively, this could be done by increasing only the peak luminance value. The consensus was to perform stretching by only increasing the peak luminance value.

It was also noted that the selection of the cd/m² value target may affect the test procedure (described below).

It was commented that JCTVC-W0046 provides a recommended conversion process between BT.709 and BT.2020 content. It was further clarified that JCTVC-W0046 does not address a change in transfer functions.

It was also suggested to use luma adjustment from JCTVC-X1020 for any conversion from 4:4:4 to 4:2:0.

It was commented that there was and/or will be discussion in other groups that considered alternative methods for luma adjustment that may be related.

One suggestion is for JVET activity using the luma adjustment tool to use the method included in future JCTVC test conditions, e.g. a hypothetical JCTVC-Y1020. This was agreed.

Recommendation of BoG: Conversion of broadcast range, legacy (SDR) content should map the broadcast legal range to the range of 100 cd/m² and 0.01 cd/m².

Recommendation of BoG: Conversion of full range, legacy (SDR) content should map the full range to the range of 100 cd/m² and 0.01 cd/m².

Recommendation of BoG: Luma adjustment from JCTVC-X1020 shall be used for any conversion from 4:4:4 to 4:2:0.

During the discussion, it was commented that several of the CTC sequences have values that are outside of the broadcast legal range. This was reported on the reflector during AhG9 activities.

It was discussed whether the conversion procedures should address this issue. However, as a result of this discussion, it was decided to raise this issue to the JVET with the following recommendation:

Recommendation of BoG: The BoG recognizes that some of the CTC content has values that exceed the legal range, and recommended that JVET adjust the CTC content so that the content is in the legal range.

Note that after review by JVET, the BoG was tasked with determining a way forward to adjusting the CTC content. Discussion about broadcast legal range and test content follows:

It was suggested that for content that has a limited number of samples out of range or the overshoot is not severe, new versions of the content be generated by clipping the values.

One suggestion was to apply clipping to all of the sequences.

It was reported that Cactus is likely full range as are Class F.

One suggestion was to determine what sequences are full range by viewing and then clip the rest of the sequences.

For the purpose of the conversion process, it was suggested that we apply clipping in the RGB domain within the target container. For example, SDR content may exceed the range of the BT.709 colour space but not the BT.2020 colour space. In this case, clipping would not be applied. This was agreed.

Recommendation of BoG: Perform visual evaluation of the sequences after clipping. If the clipping creates noticeable artefacts, then treat the content as full range. Otherwise, apply clipping and use the clipped sequences going forward. An AhG was recommended to further study this.

Test procedure:

It was suggested to use the configuration described in JCTVC-X1020 (with luma QP adaptation and chroma QP adaptation) for the anchor.

It was noted that the table used for luma QP adaptation likely needs to be changed for the SDR originating content and/or new content. One participant reported that it would be possible to derive the QP adaptation table for SDR content explicitly given the conversion process.

It was also noted that the chroma QP adaptation may need to account for the QP+1 adjustment already used in the JVET CTC and may also need study for SDR and/or new content.

The group also discussed the selection of the QP parameters for testing. Following the AhG9 discussion, it was suggested to use the term “master QP” when discussion the test, as the QP value may be varying based on luma and chroma and so is no longer a “fixed QP”.

It was suggested that finer QP values may be needed for testing high dynamic range and wider colour volume content.

For the master QP – one suggestion was to use the QP values used in the current JVET CTC and, in addition, add an additional QP point that is finer. For example, QP 17.

Recommendation of BoG: Use the luma QP adaptation and chroma QP adaptation for the anchor.

Recommendation of BoG: Use QP values of [37, 32, 27, 22, 17]

One suggestion was to evaluate additional QP values around the coarsest QP values. This was motivated by experience with darker sequences.

It was commented that we may need different QP values for different content and/or different content classes, for example, brighter and darker content.

It was suggested that the QP points be further discussed in an AhG. This was agreed.

Performance measurement:

There was some discussion on measuring performance. Potential metrics could include tPSNR and deltaE. However, multiple participants also recommended using visual testing.

It was suggested to make the reporting of tPSNR-Y and deltaE100 as part of the CTC.

It was suggested to study PSNR-L100 and the weighted PSNR metric (proposed in JVET-D0124) in an AhG.

It was suggested to include PSNR-L100 as part of the CTC as an optional metric.

It was suggested to include PSNR in the Y, Cb, and Cr domain.

Recommendation of BoG: The CTC should include (i) PSNR in Y, Cb and Cr, (ii) weighted PSNR in Y, Cb, and Cr with a grey background, (iii) tPSNR-Y, (iv), deltaE100, and (v) PSNR-L100.

It was suggested that the metrics be further studied in an AhG.

It was suggested to include the HM as an additional anchor and task the anchor generation to an AhG. This was agreed.

It was suggested that the HM should have the same tuning as the anchor discussed above. This was agreed.

There was a discussion in JVET plenary Wed. about the violation of legal range for some test sequences – reviewing the following options:

• not using sequences that violate the range
• process content such that range becomes legal
• check with providers of content whether RGB 4:4:4 versions still exist, as the violation may be due to conversion process which could be corrected

As it needed more clarification, this issue was further discussed in the BoG. The recommendation coming out of these follow-up discussion was to apply clipping, which needs further investigation in an AHG, but may not be viable for all sequences.

The BoG reported back on Friday. The recommendations were confirmed, as applicable. Most of them required further study. The JVET-D1020 draft was also presented and approved. Further notes from this discussion are as follows.

The AHG shall develop software for content-dependent luma and chroma QP adaptation, as separate branch of JEM4.0, two weeks after the JEM4.0 release. This will be used for generating anchors in HDR/WCG experiments.

The CTC will include those available sequences that are appropriate in usage terms. New sequences will also be investigated.

For the new test material (Netflix), 25 sequences were pre-selected, and to assess their characteristics in terms of compression, HM (not JEM) encoding in the “common practices” style should be conducted. This should be included in the test work plan JVET-D1002 (update on this aspect 1 week after the first delivery). A. Norkin and A. Segall were asked to identify volunteers, and J. Samuelsson was asked to give assistance in software practices.

EE6: Adaptive scaling for HDR/WCG material

This includes JVET-D0118 and JVET-D0124, which derive the luma-adaptive quantizer scaling factor at the decoder. These will be compared against the HDR anchors. Subjective testing was also planned.