Open Chris-city opened 1 year ago
For this paper, the output of offset_conv
is initialized to be 0 and does not have constraint during training.
In theory the offset can be very large but in practice it usually attends more on semantic meaningful regions.
Thank you for your work. I have a question, whether the range of offset_conv related to the kernel size of offset_conv, for example, whether the offset_conv of 3*3 can only consider the offset of one point around each pixel in the layer?