JVET-X0064 EE1-2.2: CNN-based Super Resolution for Video Coding Using Decoded Information [C. Lin, Y. Li, K. Zhang, L. Zhang (Bytedance)]
This contribution reports the EE1-2.2 test results of JVET-W0099. JVET-W0099 presents a convolutional neural network-based super resolution scheme for video coding. Compared with VTM-11.0 with new MCTF enabled, the proposed method reportedly shows on average {-6.15%, 5.58%, -2.31%} and {-10.04%, 20.44%, -6.16%} BD-rate changes for {Y, U, V} under RA and AI configurations, respectively.
Luma: Two different models for I-slice and B-slice (I-slice model larger); chroma only one model.
Is the better performance in intra due to the larger model with 96 channels? The proponent mentions that the performance does not drop much when using 64 channels.
Why only L1 as training criterion? No specific reason (except that it is commonly used in SR), L2 fine-tuning might improve results
Processing is currently done in 128x128 blocks. Would line-wise processing be possible?