JVET-AG0062 Non-EE2: Multiple Transform Sets Selection for LFNST/NSPT [F. Wang, Y. Yu, H. Yu, D. Wang (OPPO)]

This contribution proposes a multiple transform set selection method for intra blocks coded with LFNST/NSPT. In the proposed method, CUs coded with DIMD, TIMD, MIP, SGPM and ITMP modes can choose an alternative LFNST/NSPT transform set. The transform set selection is signaled in the bit-stream. On top of ECM-11.0, the performance of the proposed method is reported as follows:

AI {-0.08% Y, 0.06% U, 0.05% V, 107.9% EncT, 100.7% DecT}

Complete RA results were not available at the time of presentation. Class B results were reported to be {-0.06% Y, 0.10% U, 0.05% V, 106.3% EncT, 100.0% DecT}.

The proponent asserted that the encoding runtime increase was mainly due to unoptimized encoder implementation.

The proposal introduces an additional CU-level flag for CUs coded with DIMD, TIMD, MIP, SGPM and ITMP modes. It was asked if there are other conditions that determine whether this flag is signaled that might introduce parsing dependency, the proponent reported that no parsing dependency would be introduced.

When the two intra prediction modes used to select the transform set index are very similar (i.e. less than or equal to 4), what will happen? The proponent asserted that the usage of DIMD derived mode could guarantee that the second mode to be sufficiently different from the first.

Transform kernels are not modified by this proposal.

It was commented that selecting from different transforms usually requires full RDO, and therefore the encoder runtime increase could be difficult to bring down.

It was commented that the idea in this contribution was interesting, however the current tradeoff did not seem to be promising.

However, multiple experts expressed interest in studying this in the next round of EE.

It was agreed to investigate this in the EE. A reduction in encoding time without compromising the gain would be needed for this method to be considered valuable.