Thank you for your work.Could you explain why in the final step of deformable convolution p = self._get_p(offset, dtype) is followed by p = self.opt(p) convolution? It seems like the previous steps are useless this way. Wouldn't it be better to use self.opt from the beginning to derive the offset?
Thank you for your work.Could you explain why in the final step of deformable convolution
p = self._get_p(offset, dtype)is followed byp = self.opt(p)convolution? It seems like the previous steps are useless this way. Wouldn't it be better to useself.optfrom the beginning to derive the offset?