Hi Stefanie,

Wow, those results *do* look weird! =) Why are they not the same shape as the image? Could you send us the complete code you used?

In the meantime, there's a few things to change. `cut_threshold` is probably the single most fragile algorithm you can use with a RAG. You want to be using merge_hierarchical. You can do this either on a "color graph", which merges regions according to color similarity (and is most similar to the statistical region merging of Fiji):

or on a *boundary graph*, which first detects "edges" in the image and then merges regions progressively according to the mean edge value:

My intuition about this image is that your best choice is to use a sobel filter:

followed by compact watershed:

then by region boundary merging:

The edges look pretty sharp, too. You might even get good results with Canny:

On 15 November 2016 at 7:11:14 pm, Stefanie Lück (luecks@gmail.com) wrote:

Dear all,

sorry about the worse problem description! I tried this example with different parameters for SLIC and graph.cut_threshold  but none of them gave me satisfying results.

I attached the original image, the ImageJ SRM (Q=25) output and the RAG output (segmentation.slic(img, compactness=20; n_segments=400), graph.cut_threshold(labels1, g, 10))

The results are quite different, obviously I am doing something wrong. My aim is to segment each leaf separately. At the moment I am using felzenszwalb, which gives me quite reasonable results. However I would like to try everything possible and therefore I would appreciate some tips.

Thank you for the explanation of the algorithm, that was helpful.

Best regards,

Stefanie 

2016-11-15 1:51 GMT+01:00 Juan Nunez-Iglesias <jni.soma@gmail.com>:

Hi Stefanie,

Sorry, these responses should all be CCd to the list, so that others can benefit from the discussion — my bad for dropping that thread. Could you please:

- provide an example segmentation where skimage is doing worse than Fiji

- provide the script and parameter settings for both

Then we can help troubleshoot. I don’t know what you mean by the results “were very strange”, for example, so it’s hard to diagnose the problem. =)

Starting to merge directly from pixels, as the Fiji plugin does, is expensive and can be error prone. SLIC is a fast, initial pixel merging step, from which we can merge regions according to various criteria.  With the right parameters, SLIC + RAG mean color agglomeration should give quite similar results to Fiji’s approach…

Juan.

On 15 November 2016 at 2:39:45 am, Stefanie Lück (luecks@gmail.com) wrote:

Hi Juan,

thank you for your reply! I have seen and tested the RAG examples but I did not understand the SLIC step and the results were very strange... Is there any advantage? I am using SLIC anyway at the moment but the statistical region merging of ImageJ gives me better results.

Thanks

Stefanie

2016-11-14 13:44 GMT+01:00 Juan Nunez-Iglesias <jni.soma@gmail.com>:

Hi Stefanie!

Have a look at skimage.future.graph! There are some relevant examples in the gallery, too:

http://scikit-image.org/docs/dev/auto_examples/#segmentation-of-objects

The future.graph API is still experimental (that’s why it’s in “future”), so we really appreciate any feedback you have about it!

Juan.

 

On 14 November 2016 at 8:21:12 pm, Stefanie Lück (luecks@gmail.com) wrote:

Hi!

I am looking for a statistical region merging segmentation. Is there anything like this in skimage?

Thanks in advance,

Stefanie

_______________________________________________
scikit-image mailing list
scikit-image@python.org
https://mail.python.org/mailman/listinfo/scikit-image