JVET-R0144 AHG14: On lossless operation with RRC [J. Gan, C. Rosewarne (Canon)]
This contribution asserts that “slice_ts_residual_coding_disabled_flag” introduces a bug for mixed lossy/lossless coding when either of the dependent quantization or sign data hiding coding tools are also enabled. It is proposed to fix this bug for each of the DQ and SDH coding tools by additionally conditioning their application on the transform_skip_flag.
The proposal disables DQ or SDH locally, whenever TS/RRC is used. It still uses the context derivation of DQ, i.e. basically it aligns the text with the software for the DQ aspect; for SDH, also the software is changed.
It is further claimed that the necessary block-level change is minimum, as it is only necessary once per block and not once per coefficient (as the syntax would suggest the latter).
It is claimed that the benefit would be for mixed lossy/lossless coding, where otherwise by disabling SDH/DQ compression penalty would occur for lossy blocks.
One problem to make a judgment based on a realistic lossy/lossless use case scenario is not possible.
The whole problem is not about conceptual wrong design, but about coding efficiency.
JVET-R0110 is a software contribution that reports on experiments with mixed lossy/lossless coding. This uses lossless coding for a smaller area (slice) and reports losses in the range of 4-7% for natural content, when RRC is not used for lossless coding.
It is pointed out that, as the lossy part is in normal QP range, and therefore, the highest amount of bits likely goes into lossless part, the loss might be much lower if TS/RRC would stay enabled, but DQ/SDH disabled.
It is further pointed out that potentially a mechanism for enabling/disabling DQ/SDH at slice level would also solve the problem of better compression performance for lossy/lossless coding. Contributions JVET-R0271 and JVET-R0155 are suggesting approaches which move the control of DQ/SDH to slice level.
It is pointed out during the discussion that overriding the picture level flags for DQ/SDH when TS/RRC is enabled would be an alternative solution achieving this, as there could be arguments that the overhead at slice header would be too large.
The issue was resolved - see further notes under JVET-R0271.