关于Census Loss的实现方法

[PurpleHjl]

请问训练中的Census Loss应该如何实现？

[ltkong218]

You can refer to IFRNet.

[PurpleHjl]

Thank you for your reply.
However, while trying to reproduce the training results of SAFNet, I still met some problems causing my reproduced model to have a PSNR-mu on the SIGGRAPH2017 dataset about 0.5dB lower than the paper.
Therefore, I would like to continue to ask you some training details. Firstly, the number of training samples, according to my understanding, each epoch will use 74 training data on SIGGRAPH2017 for data augmentation before training, I would like to ask you the number of training samples in each epoch. Secondly, for the data enhancement approach in the paper I used the data enhancement approach with enhancement probabilities from your previous IFRNet, i.e.: rotation, flipping, channel reversing, cropping, etc., and I would like to ask if the specifics of the data enhancement approach and probabilities you used are significantly different from the IFRNet approach.

[ltkong218]

Your understandings are mostly right. You can try diverse data enhancement approaches.
Thanks.