How to train deformable convolutions?
There is a concept in ML called deformable convolutions, instead of kernelling over rectangle filter, we use kernelling over learned shape.
Whilst classic convolution is y(p_0) = sum(w(p_i) * x(p_0 + p_i)), deformable is y(p_0) = sum(w(p_i) * x(p_0 + p_i + d_i)).
How to differentiate over d_i if it's a shift in coordinates but not in our space?
Topic convolutional-neural-network deep-learning machine-learning
Category Data Science