JVET-AK0184 EE2-related: Additional results for NN-based ILF in ALF [D. Rusanovskyy, Y. Li, M. Karczewicz (Qualcomm), J. Li, Y. Li, C. Lin, K. Zhang, L. Zhang (Bytedance)] [late]

Contribution JVET-AJ0183 reports EE2-4.8 Test B results, where joint YCbCr VLOP filter of NNVC was tested for luma samples filtering in ECM. In this contribution, Test B results for NNVC LOP2 and HOP3 are reported to provide an initial assessment of the potential gain achievable by more complex NNVC ILF in ECM.

NN-ILF LOP and HOP models shared by JVET NNVC (JVET-AH0014) were used in this test, with no retraining over the ECM data set.

Preliminary comparative results (BD-rate change) vs ECM15 anchor are shown below.

Test B LOP (Partial results for completed sequences B*, C, D, E):
AI: -1.4%, 0.0%, 0.0% (Y, Cb, Cr) with 101% EncT and 433% DecT.
RA: -1.4%, -0.0%, -0.0% (Y, Cb, Cr) with 101% EncT and x% DecT.

Test B HOP (Partial results for completed sequences of B*, C, D, E):
AI: -4.3%, 0.0%, 0.0% (Y, Cb, Cr) with 107% EncT and 5554% DecT.
RA: -4.7%, -0.2%, -0.2% (Y, Cb, Cr) with 107% EncT and x% DecT.

The results could be interpreted to indicate that at most half of the luma gain obtained in NNVC may still be obtained with ECM. This view might change if also chroma filtering would be considered, which is known to perform better in NNLF, and strategy for luma/chroma optimization is different for ECM and NNVC.

Results for RA from encoder loop – currently there was a mismatch between encoder and decoder.

It was agreed to include this in the study for developing an interface for NNLF in ECM (see notes under EE2-4.8).