skimage.feature

[Vighnesh Birodkar]

Hello Zsuzsanna 


1. According to [1], there is a constant factor difference between DoG and 
LoG which is ignored while computing maximas. My guess is that you will be 
able to get similar results using different thresholds for both methods. 

2. Regarding accuracy, I am not sure you can get perfect results with these 
methods, since they do not go about counting individual pixels. The 
blob_log method uses gaussian blurring internally, and my guess is, 
interference from surrounding blobs is causing the difference. For you 
synthetic data you can use regionprops. And unless you real data has 
perfectly circular and single colored blobs, your measurements are going to 
be off anyways.  

3. There is not way of doing that now. How would you define the best 
detection ?

If accuracy is critical, we might be able to suggest other approaches, 
could you upload a sample of your real data set ? Or perhaps tell us what 
the images are like ?

[1] : https://www.cs.ubc.ca/~lowe/papers/ijcv04.pdf

Thanks
Vighnesh



On Friday, May 29, 2015 at 12:15:20 AM UTC+5:30, Zsuzsanna Püspöki wrote:
>
> Thanks for your message. 
>
> Yes, I have tried both. I can detect all the blobs, but somehow the 
> algorithm does not capture the size perfectly. I am wondering why there is 
> such a difference.
> Actually, my intention is to use the results provided by this python code 
> to make a comparison on real data between different algorithms (including 
> mine). But I am not very convinced by those detections, since they do not 
> meet with what I expect. 
>
> Regards,
>
> Zsuzsanna
>
> 2015. május 28., csütörtök 11:15:09 UTC+2 időpontban Kai Wiechen a 
> következőt írta:
>>
>> Have you tried to modify the overlap or the max_sigma parameters? There 
>> is an internal function _prune_blobs to remove overlapping signals favoring 
>> larger ones.
>>
>> Regards,
>>
>> Kai
>>
>>
>>
>> Am Donnerstag, 28. Mai 2015 08:35:08 UTC+2 schrieb Zsuzsanna Püspöki:
>>>
>>> I am trying to use the blob_dog, blob_log, blob_doh from skimage.feature 
>>> on some synthetic data.
>>>
>>> 1. Why do you obtain such different results for the LoG and DoG on your 
>>> sample data: 
>>> http://scikit-image.org/docs/dev/auto_examples/plot_blob.html 
>>> <https://ewa.epfl.ch/owa/redir.aspx?C=9lxjKPMqzkiQgZDrW5vS3JnRU_KdbNII73D0phYBj9Do6bHbjT_Oecmyv0YkaA_xkC-zzPeoTjQ.&URL=http%3a%2f%2fscikit-image.org%2fdocs%2fdev%2fauto_examples%2fplot_blob.html> 
>>> ? 
>>> I would expect something very similar.  
>>>
>>> 2. I am running the following line on the attached image: blobs_log = 
>>> blob_log(image_gray, overlap = 0, min_sigma=3, max_sigma=25, num_sigma=100, 
>>> threshold=.1)
>>>
>>> I cannot arrive to detect precisely the size of the bigger spots (though 
>>> the range I put is enough, hence the huge spot). The increase in num_sigma 
>>> does not solve the issue. I would expect perfect results here.  
>>>
>>> 3. Is there an easy way to visualize the N best detections?
>>>
>>>
>>>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mail.python.org/pipermail/scikit-image/attachments/20150529/745a1de9/attachment.html>