deep-linear-shapes

PyTorch implementation of "Representing Shape Collections with Alignment-Aware Linear Models" paper.

If you find this code useful in your research, please cite:

@misc{loiseau2021representing,
      title={Representing Shape Collections with Alignment-Aware Linear Models}, 
      author={Romain Loiseau and Tom Monnier and Mathieu Aubry and Loïc Landrieu},
      year={2021},
      eprint={2109.01605},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}

paper | webpage

Installation 🌟

1. Create and activate conda environment

conda env create -f environment.yml
conda activate dls

Optional: some monitoring routines are implemented with tensorboard.

Note: this implementation uses torch-points3d and an adapted version of their datasets.

2. Update submodules to use Chamfer Distance

Our implementation uses an efficient GPU implementation of Chamfer Distance to speed-up training taken from this repository and used as a submodule. To properly use this implementation run:

git submodule init
git submodule update

3. Download non-torch-points3d datasets

Execute bash scripts in the folder dlm/datasets/AtlasNetData. These files are taken from this repository.

How to use 🚀

1. Train a model

python run.py

Results are saved at runs/. Some options can be found in the file dlm/utils/parser or by running the command python run.py -h. Some training visual results like prototype evolutions and transformation prediction examples will be saved in a tensorboard file. Here is an example of learned prototypes on the ABC dataset:

2. Notebooks

Some explanatory notebooks are available in the folder notebooks/.