JVET-P0365 AHG15: Scaling matrices for LFNST-coded blocks [A. K. Ramasubramonian, G. Van der Auwera, V. Seregin, H. E. Egilmez, M. Karczewicz (Qualcomm)]
This document proposes a modified method of applying scaling matrices for LFNST-coded blocks. It is asserted that the current VVC design of applying the scaling matrices has some shortcomings. For all LFNST-coded blocks, it is observed that the LFNST coefficients are restricted to be contained in a 4x4 region of the transform block. It is proposed that when scaling matrices are applied to an LFNST-coded block, the scaling matrix corresponding to the 4x4 transform block size be applied to the LFNST coefficients. For applications that use quantization matrices for rate control, this method is asserted to be sufficient. For use of scaling matrices for perceptual quantization, it is not evident that the scaling matrices that are specified for the primary matrices may be useful for LFNST coefficients. Until such evidence is available, it is proposed that flexibility be provided to disable the scaling matrices for LFNST coefficients.
There is no evidence that the usage of 4x4 scaling matrices for all block sizes of LFNST is beneficial. Actually, the properties of LFNST coefficients may be closer to the primary transform.
It is agreed that the aspect of just disabling the scaling by an APS flag is straightforward and allows an encoder to just ignore LFNST in the design of scaling, or by purpose disabling the scaling in case of LFNST if it would produce artefacts.
Decision: Adopt JVET-P0365 (only the aspect of scaling_matrix_for_lfnst_disabling_flag in APS for scaling matrices).