CGAL
Mesh 3 - fix construction of weights image for labeled image input
Summary of Changes
Computation of weights for labeled image input was broken. Using itkDiscreteGaussianImageFilter with voxel-number standard deviation (instead of itkRecursiveGaussianImageFilter with absolute standard deviation) fixes it.
Release Management
- Affected package(s): Mesh_3
- License and copyright ownership: unchanged
Todo
- [ ] make screenshots before/after
- [ ] check that user code has the same improvement
Let's reintroduce sigma to have this PR as a real bug-fix, and we'll see later if we remove it
@sloriot This pull-request is ready to be tested, code-wise, even if there are still TODO items about subjects external to Github (communication and checks in an external user-code).
Making more experiments on some user data, I found out that there is a trade-off in the choice of parameter sigma :
- if too small we get the "stair-artefact" back
- if too large the fine volumes (volumes made by 2 close surfaces) are lost and not represented in the output mesh
So let's keep the parameter for now, though it would be easier to use if it were a number of voxels than a real-world distance
Making more experiments on some user data, I found out that there is a trade-off in the choice of parameter
sigma:* if too small we get the "stair-artefact" back * if too large the fine volumes (volumes made by 2 close surfaces) are lost and not represented in the output meshSo let's keep the parameter for now, though it would be easier to use if it were a number of voxels than a real-world distance
What is the right value, according to your experiments? About the width of one voxel?
About the size of one voxel yes. But experiments did not really bring conclusions on weather we should use the minimal dimension, or maximal dimension, for the default value (?)
@janetournois You said in the TODO list in the first message (https://github.com/CGAL/cgal/pull/6864#issue-1379697365):
make screenshots before/after