Unable to realise the effect of the article

[A1475465724]

Tested using GOPRO and RED data, set up and trained according to the parameters of the paper, the final results show that the effect in the paper cannot be achieved

[ZhijingS]

May I ask how do you split the image set B and S in the paper? Do the scenes in image sets B and S not overlap at all, or will images of the same scene be partly in B and partly in S？Looking forward your reply.

[A1475465724]

sure, For train ,we prepared four datasets. TrainA represents the blurry images from unkonwn kernel(set Beta in the paper) ,Train B represents the sharp images from the same imaging device of train A (set S in the paper). The two datasets do not need to be paired, just make sure that it is a blurry image from an unknown blur kernel in trainA and clear images from the same device in trainB. It does not matter whether the scenes in these two datasets overlap or not> May I ask how do you split the image set B and S in the paper? Do the scenes in image sets B and S not overlap at all, or will images of the same scene be partly in B and partly in S？Looking forward your reply.

[ZhijingS]

Thanks for your reply! I think if the scenes in B and S overlap, it may be easier for the blur translater to learn how to transfer blur kernel and may get better results. However I haven't try it.

[A1475465724]

Thanks for your reply! I think if the scenes in B and S overlap, it may be easier for the blur translater to learn how to transfer blur kernel and may get better results. However I haven't try it.

Thank you for your advice，Have you try this paper's code? what is your result? Does this article's theory really work?

[ZhijingS]

No, I haven't tried the code. I just take the idea of the paper and try to make a unsupervised deblur algorithm. Sadly, my experiments didn't work.

[A1475465724]

No, I haven't tried the code. I just take the idea of the paper and try to make a unsupervised deblur algorithm. Sadly, my experiments didn't work.

I have tried this paper's code.However , the results showed it didn‘t work,Maybe you can try other ways

[ZhijingS]

Thanks for your advice. Do you focus on unsupervised deblur algorithms? I think most of them need a GAN to try to create more realistic blur and then we can deblur, however the lack of a good metric for realistic blur makes it difficult to guide GAN training.