Implementation of the NeurIPS 2020 paper (arXiv). For a given pair of deformable 3D shapes, our algorithm produces high-quality correspondences.

• To preprocess the raw datasets, run the matlab file:

preprocess_data/preprocess_dataset.m

• In our experiments, we use the FAUST remeshed and SCAPE remeshed benchmarks. Both datasets can be downloaded from here.
• Change dataset paths under the get_faustremeshed_file() and get_scaperemeshed_file() functions in data.py.
• To jointly train multiple datasets (including inter-dataset pairs), create a hybrid dataset, see e.g. FaustScapeRemeshedTrain in data.py. In the paper, we show an experiment where we jointly train on FAUST remeshed and SCAPE remeshed (see Table 1).

• Checkpoints are saved under:

./models/

• The checkpoint in ./models/Faust_Scape/ corresponds to our results in the 5th and 6th column of Table 1.

• For FAUST remeshed, run train_faustremeshed_train() from main.py.
• For SCAPE remeshed, run train_scaperemeshed_train() from main.py.

• To output test correspondences for one sample pair, see demo_faust_scape() in main.py.
• Run your own examples analogously.

If you use our implementation, please cite:

@article{eisenberger2020deep,
  title={Deep Shells: Unsupervised Shape Correspondence with Optimal Transport},
  author={Eisenberger, Marvin and Toker, Aysim and Leal-Taix{\'e}, Laura and Cremers, Daniel},
  journal={Advances in Neural Information Processing Systems},
  volume={34},
  year={2020}
}

• The repo also contains a vanilla implementation of the CVPR 2020 paper smooth shells (the original implementation is in matlab).
• If you use this part of our implementation, cite the smooth shells paper:

@inproceedings{eisenberger2020smooth,
  title={Smooth Shells: Multi-Scale Shape Registration with Functional Maps},
  author={Eisenberger, Marvin and Lahner, Zorah and Cremers, Daniel},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={12265--12274},
  year={2020}
}