JVET-AN0068 AHG9: On Neural-network Post-filter Information Signalling and Semantics [S. Deshpande (Sharp)]

Following modifications are proposed related to NNPF signalling and semantics in VSEI Technologies under Consideration (TuC) document:

• Proposal 1: It is asserted that there are issues in Neural-network post-filter characteristics SEI in VSEI TuC related to the signalling of text prompt for temporal extrapolation (nnpfc_prompt[ i ]) and its semantics. Two alternative options are proposed.
  • Option 1: It is proposed to bug-fix the asserted issues and to separate the signalling of the text prompt information for temporal extrapolation purpose compared with the signalling of the text prompt information for spatial extrapolation purpose which is already in VSEI V4.
  • Option 2: Bug-fixes are proposed for the asserted issues related to the text prompt signalling for temporal extrapolation purpose.
• Proposal 2: In Neural-network post-filter characteristics SEI message an asserted bug-fix is proposed for the semantics of nnpfc_mul_intefaces_flag equal to 1, to be consistent with the derivation of NNPfInferenceProcess.
• Proposal 3: In Neural-network post-filter activation SEI message it is proposed to update a constraint as follows: When nnpfc_inband_prompt_flag is equal to 0 in the NNPFC SEI message with nnpfc_id equal to nnpfa_target_id, the value of nnpfa_prompt_update_flag[ i ], if present, shall be equal to 0. This also applies to V4.

Option 1 proposes more changes than option 2. Option 1 allows separation of temporal extrapolation spatial extrapolation. It was suggested that a single prompt could be worded to specifying both an instruction for both temporal extrapolation and for spatial extrapolation.

It was suggested to modify the option 2 loop to reference nnpfc_num_alt_instances_minus1 rather than the variable nnpfNumInferences.

Agreed for proposal 1 option 2, with the modification to loop described above.

Agreed for proposal 2.Regarding proposal 3, it was suggested that fix also applies to VSEI v4, and was considered to be the original design intent.

Agreed for proposal 3 for TuC and for VSEI v4.

JVET-AN0068 AHG9: On Neural-network Post-filter Information Signalling and Semantics

Constituent rectangles (CR)