Thank you @forrestfwilliams. Applying one convolution sounds like a good idea. The implementation change looks good to me. I am wondering what's the motivation for changing the shape of the regional filter from a disk to a donut?

Hey @yunjunz, good question! The small local_kernel is meant to represent the local deformation signal, while the regional_kernel is supposed to represent the deformation occurring outside of the local area. By not including local_kernel pixels within the regional_kernel (i.e. making the regional kernel a doughnut) we are trying to capture only the regional signal in the regional_kernel.

In practice, @dbekaert found that the regional_kernel is typically large enough that the effect of including the local_kernel pixels is negligible, but I thought it would be good to make this modification regardless. Especially since it does not require any extra steps when you're already combining the two kernels.

Thank you for the explanation @forrestfwilliams, that makes sense. Could you please add this as comments next to the code as well? I believe that would help other users too.

Can do!

Thank you @yunjunz !