Ilastik Segmentation Rendering (on Mac)

@k-dominik

ilastik 1.3.3post2

This is the raw data only with Segmentation overlay.
Could you explain what the black pixels mean?

This happens whenever there are “tied” pixels - so for a pixel the probability is 0.5 for both classes in your case.

There is an issue about in on our issue tracker: White/Transparent pixels in prediction (3rd class with 2 defined classes) · Issue #2025 · ilastik/ilastik · GitHub

What would you actually expect those pixels to behave like?

What value do you store when exporting the segmentation into a file? Are they assigned a zero?

I checked it and found, that there are two code paths that generate the segmentation - once for display (with 0s in these occurrences), once for export (doing numpy.argmax, hence favoring the first occurence of the maximum value).
With two classes you will have 1s for the tied pixels.
What would be the expected behavior?

I would say most importantly preview and what you actually save should be the same.

Otherwise I would just use round( probability ) :laughing: (insider joke)

Seriously, whatever, but I would use one of the classes and do not put it to zero.

1 Like

yeah, absolutely. We’ll discuss this internally. I kind of like adding 0s a lot, as anything else is indeed like round :smiley:

OK, fair enough, actually why not, to be honest I have not thought this through carefully. I was somehow thinking that 0 also is the “background” class, but that’s not the case, so then using 0 for undetermined could be fine.

We talked about this with the group and concluded not to introduce changes to the saved image, but instead make sure the display shows the same data :slight_smile: .

3 Likes