JVET-R0318 Alternative methods of LFNST index signalling [M. Koo, M. Salehifar, J. Lim, S. Kim (LGE)]
The adoptions on LFNST in the 17th JVET meeting had resolved complexity issues in LFNST, which led the following two features: 1) an LFNST index can be signalled in the case that Luma CBF (Coded Block Flag) is zero and Chroma coefficient distribution is compatible with LFNST in single-tree case, and 2) Chroma coefficient distribution shall be compatible with LFNST in single-tree case although LFNST is only applied to Luma in single-tree case. Here, the compatibility with LFNST indicates that coefficients should be properly zeroed-out for LFNST and at least one non-zero coefficient should exist at a position other than DC.
In this contribution, three alternative LFNST index signalling methods are proposed to remove one or all of the two features as follows:
- Method 1: Simple addition of a condition to (A) coding_unit or (B)/(C) residual_coding syntax table, in order to prevent the signalling in the case of zero Luma CBF
- Method 2: Enabling Chroma LFNST in single-tree case
- Method 3: LFNST index signalling in transform_unit syntax table
Method 1 removes the first feature, and others both of the two features. And, Method 1 and Method 2 incur minimal changes of spec, but Method 3 requires the fairly amount of spec changes. The BD-rate changes of Y/U/V are summarized as follows:
- Method 1: (A) 0.00%/0.00%/0.00% (AI) and -0.01%/-0.01%/-0.02% (RA),
(B) 0.00%/0.00%/0.00% (AI) and -0.01%/-0.01%/0.01% (RA)
(C) 0.00%/0.00%/0.00% (AI) and -0.02%/-0.03%/-0.02% (RA) - Method 2: 0.00%/0.00%/0.00% (AI) and -0.04%/-0.19%/-0.27% (RA)
- Method 3: 0.00%/0.00%/0.00% (AI), -0.02%/-0.04%/-0.06% (RA), and 0.00%/0.00%/0.00% (LD)
According to the experimental results of Method 1 and Method 3, both of the two feature make little impact on coding performance. Furthermore, Method 2 increases complexity due to the Chroma LFNST in single-tree case. It is recommended that 1) no action should be taken on current LFNST index signalling if it is thought that the two features do not incur any critical issues, or 2) Method 1 or Method 2 could be adopted if more conservative approaches are preferable, or 3) Method 3 could be considered if group is not concerned about considerable spec changes.
Similar to JVET-R0057, JVET-R0174, JVET-R0236 – see notes there.
No need for discussion, according to proponents
No action was taken on this.