This repository provides our implementation of the CVPR 2021 paper NeuroMorph. Our algorithm produces in one go, i.e., in a single feed forward pass, a smooth interpolation and point-to-point correspondences between two input 3D shapes. It is learned in a self-supervised manner from an unlabelled collection of deformable and heterogeneous shapes.

If you use our work, please cite:

@inproceedings{eisenberger2021neuromorph, 
  title={NeuroMorph: Unsupervised Shape Interpolation and Correspondence in One Go}, 
  author={Eisenberger, Marvin and Novotny, David and Kerchenbaum, Gael and Labatut, Patrick and Neverova, Natalia and Cremers, Daniel and Vedaldi, Andrea}, 
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition}, 
  pages={7473--7483}, 
  year={2021}
}

The code was tested on Python 3.8.10 with the PyTorch version 1.9.1 and CUDA 10.2. The code also requires the pytorch-geometric library (installation instructions) and matplotlib. Finally, MATLAB with the Statistics and Machine Learning Toolbox is used to pre-process ceratin datasets (we tested MATLAB versions 2019b and 2021b). The code should run on Linux, macOS and Windows.

Using Anaconda, you can install the required dependencies as follows:

conda create -n neuromorph python=3.8
conda activate neuromorph
conda install pytorch cudatoolkit=10.2 -c pytorch
conda install matplotlib
conda install pyg -c pyg -c conda-forge

In order to run NeuroMorph:

• Specify the location of datasets on your device under data_folder_<dataset> in param.py.
• To use your own data, create a new dataset in data/data.py.
• To train FAUST remeshed, run the main script main_train.py. Modify the script as needed to train on different data.

For a more detailed tutorial, see the next section.

We show below how to reproduce the experiments on the FAUST remeshed data.

You can download experimental mesh data from here from the authors of the Deep Geometric Functional Maps. Download the FAUST_r.zip file from this site, unzip it, and move the content of the directory to <NEUROMORPH_ROOT>/data/mesh/FAUST_r .

Meshes must be subsampled and remeshed (for data augmentation during training) and geodesic distance matrices must be computed before the learning code runs. For this, we use the data_preprocessing/preprocess_dataset.m MATLAB scripts (we tested V2019b and V2021b).

Start MATLAB and do the following:

cd <NEUROMORPH_ROOT>/data_preprocessing
preprocess_dataset("../data/meshes/FAUST_r/", ".off")

The result should be a list of MATLAB mesh files in a mat subfolder (e.g., data/meshes/FAUST_r/mat ), plus additional data.

If you stored the data in the directory given above, you can train the model by running:

mkdir -p data/{checkpoint,out}
python main_train.py

The trained models will be saved in a series of checkpoints at <NEUROMORPH_ROOT>/data/out/<TIME_STAMP_FAUST> . Otherwise, edit param.py to change the paths.

Upon completion, evaluate the trained model with main_test.py . Specify the checkpoint folder name <TIME_STAMP_FAUST> by running:

python main_test.py <TIME_STAMP_FAUST>

Here <TIME_STAMP_FAUST> is any of the directories saved in <NEUROMORPH_ROOT>/data/out/ . This automatically saves correspondences and interpolations on the FAUST remeshed test set to <NEUROMORPH_ROOT>/data/out/<TIME_STAMP_FAUST> . For reference, on FAUST you should expect a validation error around 0.25 after 400 epochs.

See the CONTRIBUTING file for how to help out.

NeuroMorph is MIT licensed, as described in the LICENSE file. NeuroMorph includes a few files from other open source projects, as further detailed in the same LICENSE file.