Do you have a paper to introduce this project in detail?

[houhouhouhou11]

Do you have a paper to introduce this project in detail?

[yellowYuga]

https://arxiv.org/abs/1512.02134

[houhouhouhou11]

@yellowYuga,Thank you very much . Btw, do you know the differnece of kernal and patch in this project. Where the dilation is unsed in ? kernal or patch? Thank you very much

[ClementPinard]

• kernel is the radius of area used for computing correlation. It is very similar to convolution kernel, since it's a weighted sum of the neighbourhood of a particular point, with the difference that here the sum is weighted by the corresponding neighbourhood of the second input.

• patch is the different translations we test between input1 and input2. Regular correlation is done with a translation of (0,0), and then we test different translations. for a particular translation u,v, we compute correlIation between input1[:, u:, v:] and input2[:, :-u, :-v] .

• dilation is the same as dilation in regular convolution, meaning that e.g. for a 3x3 correlation with dilation 2, you actually have 5x5 "a trous" correlation. This is not dependent to different translations tested in patch

• the corresponding dilation paremeter for patch is dilation_patch. Instead of testing every possible integer translation in [-patch_size/2, +patch_size/2] , we multiply every translation tested by dilation_patch. (Meaning we test as many different translation as with a dilation_patch of 1, but the translation are enhanced