JVET-AG0237 EE2-3.3 related: Fix on LFNST/NSPT index signalling [M. Koo, J. Zhao, J. Lim, S. Kim (LGE)] [late]

In EE2-3.3 (JVET-AG0061: Utilizing LFNST/NSPT for inter coding), it is expected that LFNST/NSPT index signalling for inter coding is aligned with intra case and performed only when last non-zero coefficient of the luma component is not at the DC position. However, in current EE2-3.3 software, an LFNST/NSPT index can be signalled even when last coefficient of luma component is at DC position. In this contribution, a fix for this problem is proposed.

The BD-rate and complexity changes relative to ECM-11.0:

• RA: -0.XX%/-0.XX%/-0.XX% for Y/U/V with 10X% EncT and 10X% DecT
• LD: -0.XX%/-0.XX%/-0.XX% for Y/U/V with 1XX% EncT and 1XX% DecT.

The BD-rate and complexity changes relative to EE2-3.3:

• RA: -0.XX%/-0.XX%/-0.XX% for Y/U/V with 10X% EncT and 10X% DecT
• LD: -0.XX%/-0.XX%/-0.XX% for Y/U/V with 1XX% EncT and 1XX% DecT.

The above two LD results were obtained by enabling LFNST/NSPT for both test and anchor. It was asked if any complete results for RA are available from the proponent and two cross-checkers. Such is not the case at the time of presentation.

This proposal suggests to determine DC_cond based only on luma DC condition, and not on Cb/Cr DC condition as done in EE2-3.3. The proponent of EE2-3.3 (and another expert) commented that this suggested change seems to be a logical fix.

In a later discussion, completed simulation results were provided as follows:

The BD-rate and complexity changes relative to ECM-11.0:

• RA*: -0.30%/0.04%/0.11% for Y/U/V with 102% EncT and 100% DecT
• LD: -0.86%/0.06%/0.01% for Y/U/V with 104% EncT and 101% DecT.

The BD-rate and complexity changes relative to EE2-3.3:

• RA*: 0.01%/-0.19%/0.02% for Y/U/V with 100% EncT and 100% DecT
• LD: 0.01%/-0.46%/-0.35% for Y/U/V with 100% EncT and 100% DecT.

The simulation results (* = QP 27 of ParkRunning not complete yet and copied from anchor) show some coding loss of 0.02% in class A1 and class B, and coding gain of 0.01% in class C.

It was commented by one expert that the proposed change is a simplification compared to the adopted method in EE2-3.3. However, the proponent asserted that this is not a simplification, and that it rather targets redundant signaling.

As supplemental results, the proponent reported BD-rate and complexity changes relative to EE2-3.3 + JVET-AG0208 (both of which had been adopted) to be {Y, U, V: -0.01%, -0.08%, -0.10%} and {EncT, DecT: 100%, 100%}. However, given this result was not cross checked and could not be cross checked before this meeting ends, it was suggested to investigate this in EE.

It was agreed to investigate this in the next round of the EE.