So, we definitely are not wanting to re-invent the wheel, unless there is a clear reason why our wheel is better the previous one. In short, most of the scikit is focused on the manipulation of 2D images, whereas scipy.ndimage works on ndimages. So if you can code an equivalent algorithm, that is leaps and bounds faster or more efficient than one in scipy.ndimage, I would be all ears. But I definitely don't want to get mired down in replicated code that doesn't serve a useful purpose beyond what is already in numpy/scipy. These comments should not in any way be construed as me passing judgement on your code (I havent really had a chance to go through it yet). I just wanted to make sure we are all on the same page. Will try to go through this new code this weekend. Cheers! Chris On Tue, Sep 21, 2010 at 3:57 PM, Emmanuelle Gouillart < emmanuelle.gouillart@normalesup.org> wrote:

...

...

I did not have the time to read thoroughly the code of your median filter, I think you wanted to know if this was efficient Cython code or not. I'll do it soon, maybe after you write a test function for this filter? By the way, what if the advantage of this function compared to scipy.ndimage.median_filter? No there is no reason to use our own rather than scipy's one (except that my implementation might be less obfuscated because it is less complicated) it was more an example of what we could do to bind C and python

Well, clearer and simpler code is definitely an advantage :-)