I tried to use “pixel classification” to do some segmentation, but when I choose “feature selection” after labelling, ilastik always shows “not responding” even if I change to a computer with high-grade configuration. Could anyone please help me about that?
that you experience these periods of ilastik “not responding” is really unfortunate. I tried reproducing it with the current version, but had no luck doing so. Could you maybe add a few details on your setup, please? Particularly interesting would be the version of ilastik you are using, and the type and version of your operating system (linux, mac, windows).
And just to add a little detail, this is how I tried to reproduce the issue:
- Opened ilastik, pixel classification.
- Added an image file
- Selected some features
- Added annotations for two classes in live-update mode.
- Disabled live-update mode in order to be able to switch back to the feature selection applet
- Switched back there, clicked on select features, selected a different subset and said ok.
The version of ilastik is 1.3.2 and my laptop operating system is windows 10. I tried to reduce the size of my image and also reduce the annotations on my image, then “feature selection” gives the output after 30-40 minutes. Besides, the original size of my image is 3.25MB, and I had to wait 2-3 hours when I turned on live-update mode with my original image. Could you please tell me should I reduce the size of my image only, or the reduced amount of annotations is also necessary, if I need to process my work?
the times you are reporting sound very weird, given the small size of the image (especially since you have confirmed the same behavior on a bigger machine). Do you have very “dense” annotations (is a lot of your image covered by annotations?). In general you only need sparse (=few) annotations to train ilastik.
Yes. The whole image is covered by annotations to avoid noise. I tried to use some annotations but I could not remove noise from my results. So I have to made annotations densely.
The random forest is really not the ideal partner for dense annotations, as you have noticed, it gets very slow. There is not much you can do about it. Have you looked at the pixel classification video tutorial we have on youtube?
Can you maybe post your original image here (or send it to email@example.com) and I’ll have a look.
HeartDatasetS10000.tif (3.2 MB)
This one, thanks a lot.
and you care for the dark roundish objects?
For your data I would suggest the following:
- Use the autocontext workflow (=2x pixel classification):
- In the first stage go for as many classes as you have in your image (4?). Don’t overdo the annotations, it doesn’t have to be perfect
- In the second stage concentrate on your object of interest vs background segmentation (so 2 classes). Try to get the shapes you want as good as possible, but don’t worry too much about “false positives”, those can be filtered out in:
- Object classification:
- Using the original image and the prediction map from the autocontext workflow, you can use the object classification workflow to filter out “false positives”. Video tutorial can be found here: https://youtu.be/SQeRGvHeT3o
make sure to use sensible features here (e.g. shape related features, but it depends on how your false positives look like)
The results look much better, thanks a lot!