JVET-AF0082 Non-EE2: on LFNST/NSPT for inter coding [F. Wang, J. Gan, Y. Yue, H. Yu, D. Wang (OPPO)]

This contribution proposes to utilize LFNST/NSPT for inter coding. For inter coded block, an intra prediction mode is first derived according to the inter prediction block. Then the derived intra prediction mode is used to select an LFNST/NSPT transform set and the transform can be processed with the selected LFNST/NSPT kernel, like in the intra coding process. On top of ECM-10.0, the performance of the proposed inter LFNST/NSPT is reported as follows:

RA {-0.37% Y, 0.14% U, 0.09% V, 114.0% EncT, 101.3% DecT, 100.0% VmPeak}

It was asked if this would create any parsing dependency, it was confirmed by multiple experts that this proposal doesn’t have parsing dependency issues.

It was asked if new transform kernels for inter coding are trained, the answer is no. Only signalling (of which kernel is used) is different from intra coding.

Some preliminary results of improved encoding time are shown in the uploaded PPT package.

Several experts commented that the gains are attractive, and that there is evidence that encoding time can be reduced.

It was asked if there is a block size restriction, the answer is that current implementation has no such limitation.

It was suggested to enable intra LFNST in LDB and use that as the anchor for EE test in LDB config. This was agreed.

Is this applied to luma only? Yes.

It was agreed to investigate this in an EE, with a reduction in encoding time desirable.