Dilation and Deformable Convolution

[FHellmann]

First of all, thanks for the great code you provide with this repo!

I've been able to successfully embed your code into my project and run it. However, I tried to use a different dilation for the deformable convolution layer but it didn't work quite as expected. I used the VGG-16 and replaced the convolutional layers in block Conv5 to modulated deformable convolutional layers. When changing the dilation to any different than 1 of these modulated deformable convolutional layers, the sensitivity drops to zero. Is this as expected? What is the explanation for this behavior?

Thanks for your help!