JVET-AC0135 AHG9: Separate activation of colour components for neural-network post-filter [C. Lin, Y. Li, Y.-K. Wang, K. Zhang, J. Li, L. Zhang (Bytedance)]

Currently, all colour components are controlled collectively in the neural-network post-filter activation (NNPFA) SEI message. With such a design, all of, or none of, the colour components should be processed by the neural-network post-processing filter (NNPF). However, the NNPF may degrade the performance of one colour component while improving the performance of the others in certain cases. Therefore, it is beneficial to only activate the target NNPF for certain colour components.

This contribution proposes to provide the capability of specifying activation of NNPF separately for different components (luma or chroma) in the NNPFA SEI message. It is proposed to use one of the following two options:

1. Signal an indicator to indicate which of the components present in the output tensor are used.
2. Signal four flags to indicate which of the components present in the output tensor are used.