JVET-Q0473 AHG15: Demultiplexing joint CbCr before dequantization [P. de Lagrange, F. Le Léannec, E. François, P. Bordes (InterDigital)]
It is proposed to move joint CbCr demultiplexing from after inverse transform to before dequantization, applying regular quantization control to each reconstructed chroma channel, so that most quantization control parameters specific to joint CbCr can be removed. This is said to simplify draft text and improve design consistency. BDRate difference using the CTCs in AI and RA is reported to be 0.00% and +0.01% respectively.
Question: Why is it a difference at all? Probably due to some rounding/clipping in dequant. Further, some optimizing encoders might want using separate quantization control to JCCR.
One drawback is that now two transforms need to be done all time.
Further study could potentially be interesting on solutions which provide an exact same result, regardless if the sign is applied before or after dequantization.
No action was taken at this point.