JVET-AG0230 EE2-3.3 related: On Inter-LFNST [S. Puri, K. Naser, C. Bonnineau, F. Le Léannec (InterDigital)] [late]

The contribution JVET-AF0082 proposes to use LFNST/NSPT for inter blocks. In this contribution, a small fix is proposed on top of JVET-AF0082 to avoid signaling LFNST index when inter block is SBT coded. Next, the concept of LFNST/NSPT is further extended to the IBC coded blocks.

Compared to ECM-11, the average BDR gains and runtimes of the extensions reported in this contribution are as follows:

-0.02% 0.00% 0.01% 101.2% 97.5% for the AI configuration,

-0.34% 0.17% 0.12% 100.5% 96.1% for the RA configuration,

-1.66% -0.15% -0.56% 111.2% 97.5% 101.2%for the Low Delay B configuration.

Since LFNST is disabled in current CTC of ECM-11, the LDB results reported above are obtained by enabling LFNST for both the anchor and test.

Compared to JVET-AG0208, the SBT related fix is exactly the same as what was proposed in JVET-AG0208.In addition, enabling LFNST/NSPT for IBC coded blocks is proposed in this contribution. The proponent reported that the coding gain due to this aspect is {-0.02% 0.00% 0.01%} for {Y U V} and {101.2%, 97.5%} for {encT, decT} for the AI configuration for natural content, while the coding performance impact for RA and LDB is not readily available at the time of presentation.

As IBC was more frequently used for screen content, it was asked what is the performance impact due this IBC-related aspect on screen content. At the time of presentation, the only data available is {-0.05% -0.21% -0.08%} for {Y U V} and {100.9% 99.7%} for {encT, decT} for class F in AI config. Other data is not available. Given IBC was used more frequently in screen content, higher encoder run time increase should be expected compared to natural content. However, contradictory observation was made in the reported data.

The proponent of EE2 Test-3.3 commented that this IBC-related aspect seems interesting, and when they tried it before they were not able to obtain coding gain.

It was asked how LFNST/NSPT is handled for intraTMP coded blocks. DIMD is used to derive the intra prediction mode to determine LFNST/NSPT index. This is already done in ECM-11.

It was commented by multiple experts that the gain achievable by the IBC-related aspect is rather small, and not justified by the encoder run time increase of ~1%.

Further study of this second aspect is encouraged.