MongeNet

This repository contains the official implementation of the paper "MongeNet: Efficient sampler for geometric deep learning" published in the Conference on Computer Vision and Pattern Recognition (CVPR21).

If you find our code or paper useful, please cite

@InProceedings{Lebrat:CVPR21:MongeNet,
    author    = {Lebrat, Leo and Santa Cruz, Rodrigo and Fookes, Clinton and Salvado, Olivier},
    title     = {MongeNet: Efficient sampler for geometric deep learning},
    booktitle = {Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
    month     = {June},
    year      = {2021}
}

For further information or questions, please email leo.lebrat@csiro.au or rodrigo.santacruz@csiro.au. See below detailed usage instructions:

Installation

This software was developed using a High Computing Platform with SUSE Linux Enterprise 12, Nvidia P100 GPUs, python 3.7.2, CUDA 10.1.168, CUDNN v7.6.4, and Virtualenv to manage python dependencies. The installation is as follows,

1. Create a python environment:

virtualenv --python=python3.7 ./mongenet_venv/
source ./mongenet_venv/bin/activate

2. Install python libraries using PIP (you can also check the requirement files generated in ./requirements.txt):

pip install torch==1.7.0+cu101 torchvision==0.8.1+cu101 torchaudio==0.7.0 -f https://download.pytorch.org/whl/torch_stable.html
pip install hydra-core --upgrade
pip install trimesh 
pip install matplotlib
pip install tensorboard

3. Install GeomLoss library and its dependencies from the official git repo see this link for the detailed instructions.

Usage

This repository is composed of python scripts parametrized by Hydra configuration files. The main scripts are in the root directory while the default configuration files for these scripts are in ./configs/ folder. Our default configuration files are densely commented explaining in detail each one of the script's parameters. These parameters can be overwritten from command line or providing an external configuration file.

Data

MongeNet is trained over a dataset of 2D triangles and their respective optimal transport based point cloud approximations using semi-discrete optimal transport. To facilitate the use of this code base, this training data is available at CSIRO-DAP.

Training MongeNet

MongeNet can be trained as follows,

python train.py trainer.output_dir=<OUT_DIR> user_config=<CONFIG_PATH>

where <OUT_DIR> is the path to the directory where snapshots and logs will be generated, while <CONFIG_PATH> is the path to a .yaml configuration file overwriting the default train configurations at ./configs/train.yaml.

Sampling meshes with MongeNet

Once MongeNet is trained, one can sample point clouds from a batch of meshes by doing,

 python sample.py sampler.input_meshes=['<INPUT_MESH1>', '<INPUT_MESH2>', ...] sampler.output_dir=<OUTPUT_DIR> sampler.output_file_suffix=<OUTPUT_SUFIX> sampler.num_sampled_points=<NUM_POINTS>

where <INPUT_MESH1> and <INPUT_MESH2> are the input meshes, <OUTPUT_DIR> is the output directory where the generated point clouds will be saved with the filename <INPUT_MESH>_<OUTPUT_SUFIX>.ply, and <NUM_POINTS> is the number of points sampled from them. Note that these configurations can also be provided using a user-defined .yaml configuration file as before. See ./configs/sample.yaml for more information about the additional parameters provided.

Acknowledgment

This research was supported by Maxwell plus