JVET-P0110 AHG15: Quantization matrices with single identifier and enhanced prediction [P. de Lagrange, F. Le Léannec, E. François, K. Naser (InterDigital)]
This contribution is based on JVET-O0223, which proposed to change quantization matrix (QM) signalling and prediction, and allow prediction across all QMs regardless of block size, by identifying a QM with a single index, transmitting QMs in decreasing block size order, and specifying prediction as either a copy or decimation process. This contribution adds a scale factor to QM prediction, and refinement with a variable-length residual than can be used for full specification, adjustment of low-frequency coefficients, or QM offset.
The amount of the text needed to support QMs is reported to be reduced by half compared to VVC draft 6, and the proposed technique is said to save more than half of the bits needed to signal example QMs from a provided test package. (note: This may not refer the newest version of text – text editor to clarify)
The proposed features: scaled prediction across all QMs and variable-size refinement, are said to enable cost-efficient adaptation of QMs at picture level.
The bit rate saving on the “test set” of matrices is reported as 54%.
Main elements:
- New method of prediction in terms of ordering the matrices of different sizes, and interleaving between inter and intra – different from the current method, the prediction starts with largest matrix.
- Residual coding on top of prediction with diagonal scan that can stop at certain frequency
- Scaling prediction between inter and intra
- Removing prediction flag (i.e. always enabling prediction
The proposal shows that with the more sophisticated coding it is possible to save approx. 4% rate when quant matrices would be sent by picture, and approx. <0.1% when changed once per I refresh period.
It is however not obvious that change per picture may not be too relevant.
Question: Do we have enough evidence about relevant usage of quant. matrices to judge the benefit and usefulness of compressing matrices?