It's 14 seconds to remove 200 seams on the 512-by-342 bench image, but I note potential optimizations in the notebook.

Right now, the seam_carve function in scikit-image accepts an energy matrix, which it then uses to generate the accumulated energy matrix. Forward energy calculates the accumulated energy matrix directly, but I calculate the energy matrix backward from it to be able to use scikit's seam_carve function.

If I write my own seam removal function or change scikit's, I think I could be able to make the speed comparable to backwards energy.

Furthermore, the accumulated energy matrix calculation step in scikit uses a double for loop, which might be able to replaced with some np stuff for a speedup for both forward and backwards energy.

I should probably raise this as an issue on GitHub... thanks!