JVET-AK0258 AHG9: Demo of real-time NNPF inference for VVC decoding on consumer laptop [J. Funnell, M. Santamaria, F. Cricri, M. M. Hannuksela, M. Pedzisz, R. Yang, S. Schwarz (Nokia)] [late]
In this demo, Nokia showcases the integration of a neural network post-processing filter (NNPF) for visual quality enhancement with Versatile Video Coding (VVC) through the Versatile Supplemental Enhancement Information (VSEI) standard. The demo verifies the possibility to run NNPF inference (complexity block-wise: 34 kMAC/pixel, number of parameters 109068) for VVC-coded 1080P video in real-time on current consumer hardware. Nokia plans to release the related software in the near future.
Demo was shown on projector. The residual generated by the filter was also demonstrated, which showed an effect that edge sharpening happens.
It was asked if it would also be beneficial to use the same filter in the loop? Not investigated.
It was commented that it is interesting to see that an NNPF with double complexity of LOP can be run in real time on NPU for 1080P25 (1080P30 might be current limit according to presenter). This would mean that 1080P60 might roughly be possible with LOP. VVC uses the hardware of the Intel processor in the PC.