What is the exact use of the features in hte boundary-based segmentation?

Hello everyone,

I am currently working on particles that have defects that I need to delineate. However I am not 100% sure to understand what is the utility of the features that I need to select and how it works. Should I select features in the “Raw Data” columns or the “Probabilities” ones ? Should I select “edgeregion”, " standard (edge)" or “standard (sp)” ?

After the boundary-based segmentation I am supposed to do a classification with this segmentation and the map of the particles as the atlas so that every defect will be measured and assigned to its corresponding particle. What features should I select for this purpose and what does it change if I select other features ? :slight_smile:

PS : What exactly is a “groundtruth segmentation” ? How do you do it ?

Sorry that’s a lot of questions :sweat_smile: have a nice day !

Hello @Antoine_Cure,

this dialog… this dialog remained at the development stage for some reason. We are saying for years now, that it should be reworked. The large amount of choices is there because it was possible, not because it’s sensible. So in practice it depends also a bit on the data which features to select.

(sp) stands for superpixels, so the features are computed on the two superpixels that share the edge.
(edge) stands for edge, and those features are calculated using the pixels to both sides of the edge.

I don’t have any good advice on which features to use. I usually go with the standard edge features. Maybe @constantinpape can help here.

The groundtruth segmentation can be used to train the random forest classifier. It will automatically assign edges to the “keep”, or the “throwaway” class, based on them belonging to the same label in the groundtruth volume, or not. So in order to do this practically, you’d need a volume that was ideally generated with an external tool (such as paintera), where you manually merge all superpixels into their “bigger” objects. But again @constantinpape will have more insight on this.


Adding some information to the answer by @k-dominik:

It usually makes sense to select “sp” and “edge” for “Raw Data” and “edge” for “Probabilities”; for the probability map it (usually) doesn’t make much sense to extract region features corresponding to the superpixels, because the probability map should highlight only the boundaries. On the other hand it makes sense to use this information for the raw data, because meaningful intensity or texture features can be extracted.

I don’t have much to add on the ground-truth, I just want to highlight that you can use the Multicut workflow without it and generate the training labels yourself by clicking edges “on” and “off”.

1 Like

Hello @k-dominik and @constantinpape,

Thanks a lot for your answers it’s much clearer now ! :smile: