search

seung-lab / euclidean-distance-transform-3d

Euclidean distance & signed distance transform for multi-label 3D anisotropic images using marching parabolas.
GNU General Public License v3.0
239 stars 37 forks source link

feat: edt with voxel connectivity graph #28

Closed william-silversmith closed 3 years ago

william-silversmith commented 3 years ago

This enables touching voxel groups to be treated as separated, for example, because the shape gives away the nature of the object but the resolution or viewing angle is not sufficient to induce a gap.

william-silversmith commented 3 years ago

This is going to be memory intensive, but at least it's linear time. It would be possible to do this in quadratic time without the memory issue. Maybe worth exploring ways of doing it depending on how crunched for memory we are. Doing it the dumb way will probably cause a single evaluation to take on the order of an hour though.

william-silversmith commented 3 years ago

The quadratic time version literally takes hours. No good.

william-silversmith commented 3 years ago

Something funny is still going on. This is the differences between the old and new versions:

image
william-silversmith commented 3 years ago

This is a cool image. The previous image with the changes in the last commit. The level of detail is higher because small differences result in high contrast now that the major defect is resolved.

image
william-silversmith commented 3 years ago

Let's merge this so the basic capability is there. If we are able to use this in Kimimaro then eventually we'll need to implement a more efficient algorithm.