the detail about BD Degradation

[wenchen4321]

Can you please tell me the detail about BD degradation? What are the sigma and the kernel size of Gaussian kernel? Because the sigma of Gaussian kernel are different in Fsrvsr and DUF.

[YapengTian]

I used the BD resizing code from the RDN https://github.com/yulunzhang/RDN/blob/master/RDN_IR/RDN_TestCode/RDN_SR_RGB/Prepare_TestData_HR_LR.m.

[wenchen4321]

Thanks a lot! But I found the BD codes in rdn that the kernel size is odd number when scale factor is 3, and kernel size is even number when scale fator is 2,4,8. And I guess the kernel size of you BD_LR training dataset is that the sigma is 1.6 and kernel size is 12*12 if you follow the setting of RDN. Have you train your model with the kernel size of 13*13?

[YapengTian]

Yes, I used the default setting as in RDN code. I did not try 1313 during training. But I think 1313 probably will be better since we usually use an odd value for the blurring kernel. Thanks for letting me learn it! You might try it when you generate training samples. In addition, I found that the models trained with gaussian blur kernels can get better results. There are several recent iccv and cvpr papers comparing with EDVR but using gaussian kernels which are unfair experimental comparison.

[wenchen4321]

Thanks for your reminder!