Goodluck making something faster than ndimage without considerable effort (I've already tried in Cython and was 10x slower). I read the ndimage source and it goes to great lengths to make optimum use of the cpu cache by allocating line buffers etc...
What we need is to find a fast open source routine, wrap it with Cython, and package it with the scikit. I wouldn't suggest using the one from OpenCV unless we are desperate because it's implemented as a c++ filtering engine. However, they are using sse2 intrinsics and it's fast! Like 10x+ faster than ndimage.
2011/4/19 Stéfan van der Walt
<stefan@sun.ac.za>
> Apparently, neither of these routines like integer images as inputs
> (should be mentioned in the docs). Here's the output for float
> images:
To be more specific, scipy.ndimage.convolve does not upcast
appropriately when convolving integer arrays with floating point
arrays. This alone seems like a good enough reason to have our own
version, although I agree that we should use the 1D separated filters.
It doesn't look as if there is an easy way to coerce numpy.convolve to
do the job, so I guess I should write something in Cython?
Regards
Stéfan