JVET-AL0230 AhG11/AhG12: Performance of the NNVC ILF in ECM [D. Rusanovskyy, K. Panusopone, S. Hong, L. Wang (Nokia)]

See section 5.2.4.4.

JVET-AL0230 AhG11/AhG12: Performance of the NNVC ILF in ECM [D. Rusanovskyy, K. Panusopone, S. Hong, L. Wang (Nokia)]

An integration of NN-ILF into ECM targeting luma-only filtering have been proposed in JVET-AJ0210 and further studied in JVET-AK0183 and JVET-AK0184. In this contribution, implementation of JVET-AK0183 has been modified and NNVC ILF filters (VLOP, LOP and HOP) have been placed in parallel to deblocking filter and tested for joint YCbCr filtering.

The VLOP1 filter the following results are reported:

• AI: {-0.7%, -1.8% and -1.8%} and 101% of encoder and 306% decoder run time.
• RA: {-0.9%, -1.7% and -1.2%} and 101% of encoder and 284% decoder run time.

The LOP2 filter the following results are reported:

• AI: {-1.3%, -4.9% and -5.2%} and 102% of encoder and 527% decoder run time.
• RA: {-1.7%, -5.8% and -5.3%} and 102% of encoder and 468% decoder run time.

The HOP3 filter the following results are reported:

• AI: {-3.4%, -10.3% and -11.1%} and 110% of encoder and 6533% decoder run time.
• RA: {-5.2%, -13% and -12.9%} and 108% of encoder and 5646% decoder run time.

Contribution for information, no specific action requested. Also uses the interface from JVET-AL0228, and using (older) models from NNVC unmodified which were using VTM as basis for training.

It was commented that it might be interesting to report decoding run time with GPU.