You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Hi @chosj95 , MIMO-UNet is a fantastic work! THX for sharing!
However, there are some problems confusing me a lot.
During the training stage, the different size of blurry images are generated by nearest interpolation. However, the corresponding supervision of different size sharp images are generated by bilinear interpolation, rather than nearest interpolation again. Intuitively, the interpolation function should be the same.
In the paper, function 7&8 have denominators in loss function for normalization. However, the code of loss function hasn't the denominators that I have mentioned above. I take it that chances are that the model will have better performance with normalization in loss function.
I am eagerly waitting for your explanation. THX a lot!
The text was updated successfully, but these errors were encountered:
Hi @chosj95 , MIMO-UNet is a fantastic work! THX for sharing!
However, there are some problems confusing me a lot.
I am eagerly waitting for your explanation. THX a lot!
The text was updated successfully, but these errors were encountered: