Modify chroma QP mapping table

Chroma QP mapping tables are modified such that the relationship between luma PSNR an chroma PSNR is on average closer to the one observed with HM 16.20. The net effect is that bits are shifted from chroma to luma and BD-rate differences between VTM and HM are most consistent across components.

BD-rate between VTM 8.0 and HM 16.20: AI: -24.09% -31.32% -32.82% RA: -34.84% -46.33% -46.91% LB: -30.02% -42.62% -40.94% LP: -33.12% -44.28% -42.80%

BD-rate between VTM 8.0 plus this change and HM 16.20: AI: -24.99% -25.35% -26.86% RA: -35.76% -39.04% -40.03% LB: -30.56% -32.89% -30.77% LP: -33.62% -34.86% -32.56%

Plots showing relationship between luma and chroma PSNR: qpchroma-v1.docx

Overall alignment looks good. But it seems the chroma gains are very dependent on image resolution. It may be worth to visually check the smaller resolutions, especially for LDB/P.

Attachments: SDR__HM-16.19_vs_VTM-8.0_ctc.xlsm SDR__HM-16.19_vs_VTM-8.0CrQpOffset1_baseQPlessThan37.xlsm

Hi Frank, all (re-posting this in case my email reply to the JVET reflector got blocked due to the Excel sheets),

thanks very much for the information. Your proposed encoder configuration change looks good to me, but for the record, I've done some slightly different experiments myself for the same purpose of reducing the luma-chroma BD-R gain imbalance over HM. I modified the encoder configuration in such a way that, for all base QPs lower than 37, slice chroma QP offsets (for both Cb and Cr) of 1 along with a JCCR chroma QP offset of 0 are used. In other words, a former proposal of mine, M0090 from the Marrakech meeting, was reinvestigated (http://phenix.it-sudparis.eu/jvet/doc_end_user/current_document.php?id=4893) but only for the higher rates, not for base QP 37 or larger. The results are attached.

Interestingly, and somewhat unexplicably to me, for Random Access the BD-R gains over HM seem to be larger with my approach than with your approach in all three color components:

Yours:

Overall -35.76% -39.04% -40.03%

Mine:

Overall -35.85% -39.35% -40.30%

So, in terms of "performance", I slightly prefer my approach, but in terms of "simplicity", yours is clearly preferable since no changes to any cluster simulation scripts are required (with my proposal, one would have to treat the QP 37 conditions differently from the other conditions and change the SDR CTCs to note the use of "--CbQpOffset=1", "--CrQpOffset=1", "--CbCrQpOffset=0" with base QPs lower than 37).

Best,

Christian

@ksuehring Had a look at some of the lower resolution sequences in low-delay configuration. Didn't see any significant changes in quality. Note that changes to the chroma QP value are generally either 0 or +1. For luma QP values 37 and above, the chroma QP value is never increased wrt VTM8.

@crhelmrich Looks like your message made it through the reflector, so not going to duplicate my response here.

unmarked as a Work In Progress

merged

mentioned in commit 2190dcd2