Thanks for your reply, Stéfan. This is also what I was thinking. But I'm not sure whether it will still provide some performance gain by eliminating the outmost python for-loop through pixels. And if the filter is something as simple as np.std (or other well vectorized function), will it work OK? Will just calling back to python slow things down back? Thanks again! Regards Chencan On Monday, May 26, 2014 11:13:20 PM UTC+8, Stefan van der Walt wrote:

Hi Chencan,

On Mon, May 26, 2014 at 8:50 AM, Chencan QIAN <qianc...@gmail.com<javascript:>> wrote:

...

I guess the absence of such a function must be due to some unresolved technical obstacle?

The main obstacle is: how would you specify such a filter? If you provide it as a function call, then we have to go through Python (not Cython), which slows things down back to the speed that ndimage's generic filter would give you.

Regards Stéfan