search

NEUBIAS / training-resources

Resources for teaching/preparing to teach bioimage analysis
https://neubias.github.io/training-resources
Other
46 stars 21 forks source link

Distance transform definition #266

Open tischi opened 2 years ago

tischi commented 2 years ago

@dlegland @jni

While one typically says that distance transforms operate on binary images, I found that both the MorpholibJ and scipy.ndimage implementations replace all non-zero pixels with their distance to the closest 0 pixel. Thus, the only pixel value with a special meaning appears to be 0, while the others can have arbitrary values. Is it fair to teach it like this to students, i.e. the input image does not need to be strictly binary?

dlegland commented 2 years ago

My position would be to stick on the definition of distance transform: it is defined on binary images, and hence must be applied on binary images. Also be careful about the fact the some refs define distance map for foreground pixels (distance to nearest 0-pixel), whereas some refs use the inverse (distance map for background pixels, computing distance to nearest non-zero pixel)... For example, the "bwdist" function in the Image processing toolbox of Matlab uses the second definition; that may be a problem when translating workflow from one language to another...

MorphoLibJ allows its application to non-binary images, but it considers input image is a label image, and computes in a single pair of forward-backward passes the "union" of distance maps for each binary label. About this, it is likely that a future version splits the two possibilities, and that the binary-version ones throw exception when the input is not binary (that would be more consistent with the definition).

More generally, if an operator is defined for a binary image and can be applied on a non-binary on, my recommandation would be: take care! It is either a bug, or a generalisation that has to be documented.

To complete a little bit, there are extensions of distance transforms that take into account the intensities of the pixels/voxels (I do not remember now how they are called, however...).

jni commented 2 years ago

I think @dlegland you're thinking of this paper:

http://www.cs.cornell.edu/~dph/papers/dt.pdf

which is super badass. I had a student working on a scikit-image implementation but it got stuck in the Python -> Cython transition.

But, that's a bit off-topic: I agree with you @dlegland that one should probably take care with non-binary inputs. imho for teaching purposes, one should describe the distance transform as working on binary images.