JVET-P0393 On BD rate computation for tools affecting quantization [S. Keating, K. Sharman, A. Browne (Sony)]

Presented Thursday 9 Oct. Track A

In JVET-N0031 it was proposed to increase the number of points used in rate-distortion curves to improve the accuracy of BD rate computation. This paper studies LMCS, which modifies the effective quantization, as an example of a tool where the BD rate computation is affected. The use of a combined YUV metric, as proposed in JVET-N0341, is also studied for the same example.

It is pointed out that for cases where different interpolation methods in BD rate computation show different results, this might be an indicator that the interpolation is misleading. This was verified by comparing against a step-1 (every QP) RD curve, which was then even more different.

An extreme case was slide editing, where the interpolation seems to show benefit, which was almost gone withstep-1 computation.

The effect seems also to be related to cases where the rate gap between two measured QP points is large (mostly between QP22 and QP27)

It is also pointed out that for low QP (where only first 100 frames are tested) the results could be quite different from those for the whole sequence, if characteristics are changing (such as Food Market, Ritual Dance).

Very interesting results which indicate that BD rate might be misleading in certain cases.

However, at this late stage of VVC development, established metrics might better be kept.

Entropy coding (5)

Contributions in this category were discussed Wednesday 8 Oct. 1515–1715 in Track A (chaired by JRO).