Using Scipy.ndimage to speedup label2rgb

[cmarshak]

Description

I wanted to propose an application of scipy.ndimage.measurements to speed up color.label2rgb in case it might be helpful to the project.

I have been experimenting with superpixel segmentations, but found the biggest bottleneck for larger images was computing superpixel statistics, e.g. color.label2rgb as seen in this demo.

A scipy module for this type of analysis can be found in scipy.ndimage.measurements. I found this module capable of supporting most analyses and to be fast on very large images. For example, the equivalent of the lines in labels2rgb can be replaced by labeled_comprehension or mean.

I created a gist to demonstrate the speedup, albeit on the small astronaut image.

If this not deemed useful for this project, I hope the record of this issue might be helpful for others doing similar experiments with scikits wonderful superpixel library.

[soupault]

@cmarshak Thank you very much for performing the preliminary benchmark! I'd support your proposal as I find label2rgb very useful in many occasions. Would you like to go forward and submit a pull request with the proposed change?

[cmarshak]

Sure - I created it using the gist code. Thank you, @soupault.

#3344