JVET-N0341 On reporting combined YUV BD rates [F. Bossen (Sharp)]
(Also relevant to quantization and contains a normative technical proposal.)
It is proposed to report combined YUV BD rate numbers in addition to separate Y, U, and V numbers. Additionally, it is suggested to modify the relationship between luma and chroma QPs by adjusting the luma to chroma QP mapping table and setting a chroma offset to 0. It is asserted that this provides a coding efficiency benefit under the combined YUV metric.
In a first aspect of this proposal, it is suggested to include the following additional information in the reporting template summary sheet:
Combined YUV PSNR (already reported by the VTM software on a per-sequence basis) computed based on the MSE across all components
BD rate values for the combined YUV PSNR metric
It was commented that during development of HEVC range extension, the YUV PSNR value is also used, but weighting of 4:1:1 (in MSE domain) or 6:1:1 (in PSNR domain) is applied to the Y:U:V components, respectively. The current YUV PSNR calculation in VTM applies 1:1:1 weighting for the three colour components in the SSE domain. It was commented that equal weighting of Y:U:V components in the SSE domain might not match the rule-of-thumb weighting of 6:1:1 or 90:5:5 that we sometimes use.
It was suggested that further study of this topic may be necessary. And further study is encouraged.
In a second aspect of this contribution, a modified QPc table (mapping from luma QP to chroma QP) is proposed such that chroma QP adjustments kick in at a higher QP value. It is further proposed to remove the offset between luma and chroma QP values we adopted at Marrakech meeting.
It was commented that with the QP offset that we are currently using in the CTC, it’s equivalent to using a larger than luma QP for chroma in the low QP range and using a smaller than luma QP for chroma in the high QP range. This was agreed to be not the intended relationship between luma and chroma QP. It was commented that when discussing related contribution (JVET-N0221) on QPc table for HDR, it was suggested to maybe not have fixed QPc table in the spec text, and to use explicit signalling of QPc table instead.
No action was taken on this second aspect.
Further study encouraged to identify proper chroma QP offset values to use in the CTC.