Code for Canonical 3D Deformable Mapping paper: arXiv, web page.

git clone https://github.com/facebookresearch/c3dm.git
cd c3dm
conda create -n c3dm python=3.8

If you want CUDA support, please follow the instructions to install Pytorch. We ran the experiments using the module torch==1.5.1+cu101.

All other dependencies can be installed by running pip:

pip install -e .

Dependencies:

• pytorch 1.5.1
• pytorch3d
• pyyaml
• numpy
• PIL
• matplotlib
• visdom
• plotly (visualisation only)
• trimesh (only for metrics)

For evaluation, pass the config name for the dataset, e.g.:

cd c3dm
tar -xzf dataset_root.tar.gz
python ./experiment.py freicars.yaml --eval

The code should download the required data and pre-trained models.

For training from scratch, make sure there is no model in c3dm/exp_out, otherwise training will continue from it. Then run e.g.

python ./experiment.py freicars.yaml

The code is released under the MIT License.

David Novotny, Roman Shapovalov, Andrea Vedaldi. Canonical 3D Deformer Maps: Unifying parametric and non-parametric methods for dense weakly-supervised category reconstruction. NeurIPS 2020.

Bibtex:

@inproceedings{Novotny2020,
    author = {Novotny, David and Shapovalov, Roman and Vedaldi, Andrea},
    booktitle = {NeurIPS},
    title = {{Canonical 3D Deformer Maps: Unifying parametric and non-parametric methods for dense weakly-supervised category reconstruction}},
    url = {http://arxiv.org/abs/2008.12709},
    year = {2020}
}