Seems the implementation of offsets is different from the original paper?

[SUZhaoyu]

Hi, I'm trying to reimplement the deform conv in Tensorflow. I found in the original paper, the number of filters for the offsets is 2 * c, where c is the number of input channels while the output offset resolution is the same as the input. However, in this repo, the number of filters for offset is 2 * kernel_size ** 2 while the output resolution is the same as the final result, without any relationship with input channels. I understand the thoughts behind this implementation and I understand how it actually works. However, I am wondering whether this method will affect the final performance, especially when the number of input channels is large? Thank you.

[SUZhaoyu]

Btw, is that the reason why you put a variable called deformable_group to overcome this issue?