Ndimage measurements median

[thouis]

This branch adds per-object median calculation to scipy.ndimage.measurements, implemented by adding to the algorithms already present in _select().

[rgommers]

Looks fine to me. The one thing that's a little odd to me is that a list is returned instead of an array, but that choice was already made before.

[rgommers]

Committed in SHA: 341791e . Thanks!

About median_positions(), that sounds not super useful to me. It may also not be well-defined, because the median can be the average of two values in the array.

[thouis]

Thanks.

There was a bit of discussion about the list vs. array a while ago:
http://mail.scipy.org/pipermail/scipy-dev/2009-April/011817.html
But it got sidretracked into a general rewrite of measurements.py in python.

Part of that discussion was a goal of rewriting the whole thing in Cython (with the python as a step toward that). I just learned about the bottleneck project (http://berkeleyanalytics.com/bottleneck/), which might be an easier path to a faster implementation (via an optional prerequisite, though I don't know how scipy feels about such things in general).

[rgommers]

With prerequisite I guess you mean an optional import to replace scipy functions with faster ones? Sounds fine to me, but I'm not sure that it's much easier than just replacing it directly in scipy. It's also stretching the stated goals of Bottleneck quite a bit (it was meant to focus on NaN-related functions in numpy), so that will need some discussion with its authors.

Important thing is that someone is willing to work on it though, not much has happened since 2009 as far as I can tell.....

List vs array is not hugely important to be as I'm only an occasional ndimage user, it just looked odd to me.