Closed qhanson closed 6 years ago
I believe that each training pair should be created from a single clean image with different noises.
From the paper:
both the inputs and the targets are now drawn from a corrupted distribution (not necessarily the same), conditioned on the underlying, unobserved clean target yi
In addition, the example training pairs in Figure 3 and 4 of the original paper is supposed to be considered that a row is a training pair. Or should it be that a column is a training pair?
I think that a column is a training pair...
Is it possible to use training pairs (x,y) where x and y are both corrupted images of different source images?
I do not think so. It seems to be impossible.
Thanks for your reply. I agree that each training pair should be created from a single clean image with different noises.
In the original paper, they described that the network is trained using independently corrupted input and target pairs.
In addition, the example training pairs in Figure 3 and 4 of the original paper is supposed to be considered that a row is a training pair. Or should it be that a column is a training pair?
However, the implementation below always uses the corrupted images of the same source image.
Is it possible to use training pairs (x,y) where x and y are both corrupted images of different source images? To be clear, x is a corrupted image of the source image 1 while y x is a corrupted image of the source image 2.
https://github.com/yu4u/noise2noise/blob/9651afe0b13da32798aea730d16d2c448d4c0952/generator.py#L37 https://github.com/yu4u/noise2noise/blob/9651afe0b13da32798aea730d16d2c448d4c0952/generator.py#L38