MeshUDF: Fast and Differentiable Meshing of Unsigned Distance Field Networks

This is the PyTorch implementation of the ECCV 2022 paper MeshUDF. We provide one dummy pre-trained UDF network and code for demonstrating our differentiable meshing procedure of open surfaces.

The below instructions describe how to:

1. Setup the Python environment
2. Launch optimization from one piece of garment to another

Setup environment

Set up a conda environment with the right packages using:

conda env create -f conda_env.yml
conda activate meshudf

For speed purposes, our modified version of marching cubes is implemented in Cython. It is largely based on scikit-image implementation of marching cubes Lewiner. To compile the custom version for your system, please run:

cd custom_mc
python setup.py build_ext --inplace
cd ..

Launch reconstruction and optimization

The provided UDF network under trained_networks/udf_4_garments/ is an auto-decoder which was trained on 4 different items of garment (0:dress, 1:jeans, 2:sweater, 3:tshirt). In optimize_chamfer_A_to_B.py, we use our method to reconstruct garments associated to latent codes A and B. We then optimize latent code A such that its corresponding mesh is similar to the one of B. This is done by applying a 3D Chamfer loss directly on the meshes, thus demonstrating the end-to-end differentiability of our method.

For example, to launch the reconstruction and optimization from a pair of jeans to a tshirt, run:

python optimize_chamfer_A_to_B.py --experiment trained_networks/udf_4_garments --A 1 --B 3

Credits and citation

Feel free to use this code for academic work, but please cite the following:

@inproceedings{guillard2022udf,
  author = {Guillard, Benoit and Stella, Federico and Fua, Pascal},
  title = {MeshUDF: Fast and Differentiable Meshing of Unsigned Distance Field Networks},
  booktitle = {European Conference on Computer Vision},
  year = {2022}
}

This code is based on the repos of DeepSDF and scikit-image, whose authors we warmly thank.