extending the random walker algorithm to multichannel images sounds like
a very nice idea, and I'm sure that it would be useful for other people
In fact, I had started working on some improvements of the random walker
code but didn't go as far as proposing the pull request. I have a branch
for that on
you can include these changes when you contribute your changes?
On Mon, Aug 20, 2012 at 10:03:53AM -0700, Josh Warner wrote:
> I have modified the existing random walker algorithm into a fully
> backwards-compatible version which allows inclusion of multispectral data,
> e.g. RGBA channels or different (registered) image modalities. Ã¯Â¿Â½I really
> liked the existing algorithm, so I just extended it rather than write one
> from scratch for my own purposes. Ã¯Â¿Â½The overhead is minimal; multispectral
> processing is triggered if data is passed as an iterable of arrays rather
> than just an array. Ã¯Â¿Â½
> This amounts to combining image gradients as sqrt(sum-of-squares) and
> dividing by sqrt(#channels). Ã¯Â¿Â½For obvious reasons, the several channels
> must be pre-processed to have data on similar ranges by whitening or a
> similar method. Ã¯Â¿Â½Not usually a problem for RGB, but in medical imaging
> this rears its head.
> Would this be of interest to the community? Ã¯Â¿Â½I'd be happy to contribute
> the changes if there is interest.
In the context of my previous email to pythonvision
I made a timing comparison between mahotas & scikits-image.
operation | mahotas | pymorph | skimage
erode | 10.80 | 14.33 | 80.17
dilate | 11.44 | 8.93 | 41.59
open | 22.45 | 23.20 | 80.18
center mass | 7.05 | NA | NA
sobel | 75.03 | NA | 105.72
cwatershed | 201.03 | 56586.50 | 290.41
daubechies | 19.05 | NA | NA
haralick | 306.48 | NA | 7391.37
(Best viewed with fixed-width fonts)
The unit of measurement is the time it takes to run ``numpy.max(image)``
Mahotas is always faster than skimage (although pymorph is better for certain
morphological basic operations). I used GCLM in skimage to stand for Haralick,
which is a rough approximation.
I attach the script that generates these (against github skimage). If you
think that I have used skimage incorrectly, please let me know.
Luis Pedro Coelho | Institute for Molecular Medicine | http://luispedro.org
On Aug 31, 2012 3:48 PM, "Tim Sheerman-Chase" <timsc60(a)googlemail.com>
> Hi all,
> I just implemented a stand alone piecewise affine image warp function. A
friend pointed me at this project and said you might be interested. I
noticed this thread on the recently merged geometric transforms. It seems
to include an affine transform function. I wondered if it would be worth
extending the new affine transform in scikits-image to also handle
> Any thoughts?
Thanks for getting in touch; we are definitely interested! I think you'll
find it easy to integrate your code with our warp module.. All you need to
specify is a coordinate transformation.
Would you like to have a look at our code as well? We'd gladly work with
you to get this integrated.
In principle I'd be interested and if you don't find any one else, I can
The only issue is that I am super swamped. But for funding skimage, I'm sure
I could make some room ;)
On 08/29/2012 03:56 PM, StÃ¯Â¿Â½fan van der Walt wrote:
> Hi everyone
> There is an opportunity to apply for some funding, and I'd like to
> know who'd be interested to co-author the grant. I started a Google
> document for this purpose, so let me know and I will add you.
> Here's to a healthy and well-funded skimage!
There is an opportunity to apply for some funding, and I'd like to
know who'd be interested to co-author the grant. I started a Google
document for this purpose, so let me know and I will add you.
Here's to a healthy and well-funded skimage!
regionsprops only measures the properties of region labeled with
strictly positive labels (since it uses ndimage.find_objects). However,
when using skimage's label with the background keyword argument, the
background pixels are labeled with -1, so that there is one foreground
region labeled with the label 0. This region will not be taken into
account by regionprops.
This behaviour can be fixed easily in regionprops, but wouldn't
it be better to label background pixels with 0?
I do not know well the algorithm that qhul implements, and it is lucky to
be more complicated than what is needed in this case since it seems to work
in higher dimensions than 2D. Nevertheless, I do not see any reason why the
points should be pathological:
here my modifications:
@@ -51,9 +51,15 @@ def convex_hull_image(image):
'scipy >= 0.9.')
# Find the convex hull
- chull = Delaunay(coords).convex_hull
- v = coords[np.unique(chull)]
+ #chull = Delaunay(coords).convex_hull
+ #v1 = coords[np.unique(chull)]
+ import vigra
+ #print v1
+ #print v2
As you can see I just exchanged the computation from scipy with that of
another library to which I personally had contributed implementing the
standard monotone chain convex hull algorithm.
Everything seems fine now.
Il giorno venerdì 24 agosto 2012 13:43:43 UTC+2, Stefan van der Walt ha
> Hi Luca
> > I confirm the bug to be really in qhull.
> > I substituted the computation of the convex hull in scipy (qhull) with
> > another library (vigranumpy) and thins fix the problem.
> Strange, do you know if the points are pathological for some reason?
> Perhaps we can pre-filter them to prevent any crashes. Where did you
> get these coordinates from?
Nice. I guess they work a lot like SLIC, only they put more emphasis on
the (x,y) space.
On 08/26/2012 03:29 PM, StÃ¯Â¿Â½fan van der Walt wrote:
> Andreas, look at these nifty super-pixels :)
> ---------- Forwarded message ----------
> From: *Nicolas Rougier*
> Date: Sun, Aug 26, 2012 at 6:26 AM
> Subject: Re: Sprint
> Also, I forgot a probably useless filter based on voronoi cells. Just
> don't know if it might be of some interest or not for scikit-image:
> Code at: https://github.com/rougier/gallery (showcase/showcase-9.py)
Andreas, look at these nifty super-pixels :)
---------- Forwarded message ----------
From: Nicolas Rougier
Date: Sun, Aug 26, 2012 at 6:26 AM
Subject: Re: Sprint
Also, I forgot a probably useless filter based on voronoi cells. Just don't
know if it might be of some interest or not for scikit-image:
Code at: https://github.com/rougier/gallery (showcase/showcase-9.py)