Unlike its namesake, this Python 3 library does not (yet) seek to conquer Eternia but to turn meshes into skeletons.
skeletor implements a number of different skeletonization methods, including mesh contraction, edge collapse, TEASAR, wavefront propagation and mean curvature skeletons -- see documentation and benchmarks below. We also provides a number of pre-/post-processing methods to clean up in- and outputs.
Please see the changelog for a summary of recent changes.
pip3 install skeletorFor the dev version:
pip3 install git+https://github.com/navis-org/skeletor@masterAutomatically installed with pip:
networkxnumpypandasscipyscikit-learntrimeshtqdmpython-igraphncollpyde
Optional because not strictly required for the core functions but recommended:
- fastremap for sizeable speed-ups with some methods:
pip3 install fastremap - robust_laplacian for more robust Laplacian operators:
pip3 install robust_laplacian - pyglet is required by trimesh to preview meshes/skeletons in 3D:
pip3 install pyglet
Please see the documentation for details.
The change log can be found here.
For the impatient a quick example:
>>> import skeletor as sk
>>> mesh = sk.example_mesh()
>>> # To load and use your own mesh instead of the example mesh:
>>> # import trimesh as tm
>>> # mesh = tm.Trimesh(vertices, faces) # or...
>>> # mesh = tm.load_mesh('mesh.obj')
>>> fixed = sk.pre.fix_mesh(mesh, remove_disconnected=5, inplace=False)
>>> skel = sk.skeletonize.by_wavefront(fixed, waves=1, step_size=1)
>>> skel
<Skeleton(vertices=(1258, 3), edges=(1194, 2), method=wavefront)>All skeletonization methods return a Skeleton object. These are just
convenient objects to represent and inspect the results.
>>> # location of vertices (nodes)
>>> skel.vertices
array([[16744, 36720, 26407],
...,
[22076, 23217, 24472]])
>>> # child -> parent edges
>>> skel.edges
array([[ 64, 31],
...,
[1257, 1252]])
>>> # Mapping for mesh to skeleton vertex indices
>>> skel.mesh_map
array([ 157, 158, 1062, ..., 525, 474, 547])
>>> # SWC table
>>> skel.swc.head()
node_id parent_id x y z radius
0 0 -1 16744.005859 36720.058594 26407.902344 0.000000
1 1 -1 5602.751953 22266.756510 15799.991211 7.542587
2 2 -1 16442.666667 14999.978516 10887.916016 5.333333
>>> # Save SWC file
>>> skel.save_swc('skeleton.swc')If you installed pyglet (see above) you can also use trimesh's plotting capabilities to inspect the results:
>>> skel.show(mesh=True)Each panel highlights one method (data points + fit); the faint lines in the background are the fits for all the other methods, so you can compare them at a glance (note the shared, logarithmic time axis).
Benchmarks
were run on an Apple M3 Max (36 Gb memory) with the optional fastremap
dependency installed. Note some of these functions (e.g. contraction and
TEASAR/vertex cluster skeletonization) can vary a lot in speed based on
parameterization.
Pull requests are always welcome!
Mesh contraction and the edge collapse approach are based on this paper:
Au OK, Tai CL, Chu HK, Cohen-Or D, Lee TY. Skeleton extraction by mesh contraction. ACM Transactions on Graphics (TOG). 2008 Aug 1;27(3):44.
Mean curvature skeletons are based on the following paper:
Tagliasacchi A, Alhashim I, Olson M, Zhang H. Mean Curvature Skeletons. Computer Graphics Forum (SGP). 2012;31(5):1735-1744.
The wavefront approach corresponds to a Reeb graph of the geodesic distance
function on the mesh: connected level sets of the distance field are collapsed
to their centroids to form the skeleton. The core construction was described
by:
Verroust A, Lazarus F. Extracting skeletal curves from 3D scattered data. The Visual Computer. 2000;16(1):15-25.
See also the Reeb graph framing in Hilaga et al., Topology Matching for Fully Automatic Similarity Estimation of 3D Shapes, SIGGRAPH 2001 and Ge et al., Data Skeletonization via Reeb Graphs, NeurIPS 2011.
Some of the code in skeletor was modified from the Py_BL_MeshSkeletonization addon for Blender 3D created by #0K Srinivasan Ramachandran and published under GPL3.
The mesh TEASAR approach was adapted from the implementation in meshparty by Sven Dorkenwald, Casey Schneider-Mizell and Forrest Collman.

