This repository contains the code for the medial skeleton extration technique described in Neural skeleton: implicit neural representation away from the surface, Mattéo Clémot, Julie Digne, Computer & Graphics, Proceedings Shape Modeling International, 2023. Our paper was granted the Graphics Replicability Stamp and the SMI 2023 Best Paper Award.
The code is tested under Linux 22.04 and Windows 11.
We use Python 3.10 for compatibility with the bpy library. The following packages are required:
- bpy (only for renderings)
- gudhi
- matplotlib
- NumPy
- PyGEL3D
- PyTorch (>= 1.13)
- SciPy (>= 1.9)
- tabulate
To install on Linux, run the following commands:
git clone --recursive https://github.com/MClemot/SkeletonLearning/
cd SkeletonLearning
sudo apt install python3.10-venv
python3.10 -m venv skelenv
source skelenv/bin/activate
pip3 install bpy gudhi matplotlib numpy PyGEL3D scipy tabulate
pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118
The repository comes with pretrained networks, but networks can be pretrained again by calling python pretrain.py.
python main.py [input] [output] computes a neural skeleton of the given input file, that can be either a .obj mesh file that will be sampled, or a .xyz file directly containing a point cloud with normals. It outputs a .obj skeletal mesh file.
TLDR: to compute and render the skeleton of the spot shape, run python reproduce_renderings.py spot.
This produces several files:
- Results/cvsk_Sine_spot.obj : the skeleton simplicial complex
- Results/unif_points_Sine_spot.obj : point set uniformly sampled on the surface (output of the uniform sampling step)
- Results/skeletal_points_Sine_spot.obj : extracted skeletal points (output of the skeleton tracing)
- Networks/net_64_6_Sine_spot.net : trained INR
- Renderings/spot.blend : saved blender file
- Renderings/spot.png : rendering of the shape and its skeleton
python reproduce_renderings.py [shape1] [shape2] [shape3] ...computes the neural skeleton of the given shapes with our method, then renders the skeleton along the original shape using Blender. It also outputs intermediate steps and a Blender scene file that can be used to modify the rendering parameters. The shapes can be chosen among the following list: birdcage, bitore, buckminsterfullerene, fertility, hand, helice, hilbert, lamp, metatron, protein, spot, zilla. This can be used to reproduce Figure 1 and our column in Figures 4, 5, 6.
Notice that some of these experiments can take up to several hours.
python reproduce_benchmark.py 0reproduces the comparison between skeletonization methods on the benchmark shape (with several level of noise / missing parts) (Table 1)python reproduce_cube.py 0reproduces the comparison between some skeletonization methods on a cube shape, and produces slices of the obtained SDFs (Figure 3)python reproduce_fertility.py 0applies some skeletonization methods on the fertility mesh (skeletons depicted in Figure 12), and produces slices of the obtained SDFs (Figure 13)python reproduce_torus.py 0reproduces the ablation study on a torus (with several levels of noise / missing parts) (Table 2)
If you use our code, please cite the following paper:
@article{neuralskeleton,
title={Neural skeleton: implicit neural representation away from the surface},
author={Clémot, Mattéo and Digne, Julie},
journal={Computer & Graphics 2023, Proceedings SMI 2023},
year={2023}
}
This work was supported by French Agence Nationale de la Recherche - TOPACS Project - ANR-19-CE45-0015
| Shape | License | Origin |
|---|---|---|
| birdcage | CC BY-SA | Thingi10K |
| buckminsterfullerene | CC BY | Thingi10K |
| fertility | AIM@SHAPE General License for Shapes | AIM@SHAPE |
| hilbert | CC BY-SA | Thingi10K |
| lamp | CC BY | Thingi10K |
| metatron | GNU - GPL | Thingi10K |
| protein | CC BY | Thingi10K |
| spot | CC0 1.0 | Keenan Crane |
| vase | Shihao Wu | |
| zilla | CC BY | Thingi10K |