JVET-P0197 CE6-related: LFNST transform mapping [J. Lainema (Nokia) , M.-S. Chiang, C.-W. Hsu, Y.-W. Huang, S.-M. Lei (MediaTek)]

This document proposes to change the mapping of LFNST indexes to transform matrices. Two methods are tested. Method 1 is described as follows. For LFNST index 1, the proposed approach selects transform index 1 for even and 2 for odd numbered intra prediction modes. For LFNST index 2 the selection is inverted and thus transform index 1 is selected for odd and 2 for even intra prediction modes. In addition, context coding for the second bin of the lfnstIdx syntax element is re-introduced. Method 2 is similar to method 1 but uses different mapping methods for the intra prediction modes <= 34 and intra prediction modes > 34.

It is asserted that for CTC conditions the coding efficiency is practically unaffected, but a low complexity encoder can operate at 90 % encoder runtime with +0.13%~+0.14 % bitrate impact under AI conditions and at 97 % encoder runtime with +0.11 % bitrate impact under RA conditions. This is reported to be a significantly lower BD-rate penalty than what was achieved with the comparable encoder approaches in CE6-1, where the best performing encoder side approach had +0.46 %~+0.48% impact on BD-rate for AI and +0.27 % for RA.The proposal allows still selecting the two NST types, but depending on intra mode parity the signalling is toggled. Furthermore, a context coded bin is added. A fast encoder could always select the same type and would go to the solution of CE6-1. The benefit in encoding runtime versus compression loss is still the same (slightly worse) as for CE6-1. The signalling becomes more complicated.

No action was taken on this.

Similar approaches: JVET-P0215, JVET-P0349 / JVET-P0350, JVET-P0568.