JVET-AL0225 EE2-related: On In-Loop Filtering in ECM [D. Rusanovskyy, M. Santamaria, N. Le, F. Cricri, K. Panusopone, S. Hong, L. Wang, J. Lainema (Nokia), Y. Li, M. Karczewicz, J. Wang, L. Kerofsky (Qualcomm)]

An integration of NN-ILF into ECM have been originally proposed in JVET-AJ0210 and being studied in JVET-AK0183 and JVET-AK0184. This contribution presents a development of the loop filtering proposed in JVET-AK0183. Loop filtering is being constrained by encoder to minimize the complexity-aware cost function and thus reducing the encoder and decoder run-times.

Proposed method was tested with ECM16 CTC, and reportedly provides the following BD-rate PSNR {Y, Cb, Cr} changes against the respective anchor:

• AI: {-0.3%, -0.9% and -0.7%} and approximately 92.3% of encoder and 200% decoder run time.
• RA: {-0.4%, -0.6% and -0.1%} and approximately 92.8% of encoder and 151% decoder run time.

It was proposed to study this method in EE2.

The difference compared to the method in EE2-5.4 is that the inputs to the online ALF can be more flexibly selected, such that also fixed filter can be replaced by NN-ILF (frame-level switching). The input of NN-ILF is identical to EE2-5.4.

It was commented that the decoder runtime is significantly higher than for EE2-5.4. Not exactly known why – likely due to use of a different NN-ILF model and less optimum implementation.

It was asked why the encoding runtime is decreased (also in comparison with EE2-5.4)? At the encoder, the checking of fixed filter is sometimes skipped.

Interface has some aspects different from previous EE, and encoding runtime reduction is interesting – to be studied in next EE. The interface of JVET-AL0228 should be used.