I am curious if people have considered the following. Suppose there is an N X N image, say x, and you plan to apply a filter to it with kernel K. Instead of using padding, why not just work modulo N? That is compute x * K over the group Z/NZ X Z/NZ.
Hmmm, I think I can see why this question went unanswered when you originally asked it. I don’t understand what you mean. What are Z and NZ here? Convolutional filtering is done “modulo N” at least in my understanding. The point of padding is that if you don’t pad, you end up reducing the dimensions.
So clearly you need to give a bit more detail about what it is you are proposing here as an alternative to padding. If you mean using part of the filters at the edges, I think they probably already thought of that and decided that padding is a more tractable way to achieve the same thing.