Feature requests for 0.5

[Emmanuelle Gouillart]

On Thu, Dec 08, 2011 at 01:31:46AM -0800, St�fan van der Walt wrote:
> On Wed, Dec 7, 2011 at 9:10 PM, Emmanuelle Gouillart
> <emmanuelle.gouillart at nsup.org> wrote:
> > Regarding graph-cut segmentation, it's not exactly graph-cut, but I have
> > some code (purely Python) implementing the random walker segmentation
> > algorithm
> > (http://cns.bu.edu/~lgrady/Random_Walker_Image_Segmentation.html). It's a
> > very nice algorithm that overcomes some of the limitations of the
> > watershed, and it can label more than two phases. So I could clean up
> > this code a bit for skimage, I guess this could be my contribution to 0.5

> Thanks for the link; I haven't seen this method before.  Does it
> vectorize fairly well?  

The implementation that I use is based on solving a large sparse linear
system, so there is no for loop or anything like this. Does this answer
your question?

> I wonder, if we start doing seeded segmentation, whether it might be
> worth building a Qt or Matplotlib GUI along with it.  I have other uses
> for a "Paint" component too, such as modifying the FFT spectrum and
> watching the changes in place (wouldn't FFT painting be fun? :).

I've got mixed feelings about this. Having some interactive GUIs for
adjusting parameters like for thresholding, or Canny filtering, would be
great and save a lot of time. Sometimes I get quite frustrated when
typing several commands in Ipython to test different parameters, whereas
the algorithm is actually fast enough so that I could just move a handle
and see the result. I asked Eric Jones during Scipy India if he could
make a small example using traitsUI for interactive thresholding.

The question is: should these interfaces go in skimage or elsewhere? If
yes, we would have to make sure that this does not introduce additional
dependencies (only optional ones). I think it's very important that
skimage stays a toolkit, and a package that it is easy to depend on for
other projects.

Cheers,

Emmanuelle