DeepCurrents | Webpage | Paper

DeepCurrents: Learning Implicit Representations of Shapes with Boundaries
David Palmer*, Dmitriy Smirnov*, Stephanie Wang, Albert Chern, Justin Solomon

Set-up

To install the neecssary dependencies, run:

conda env create -f environment.yml
conda activate DeepCurrents

Training

To prepare the training dataset, first download and extract the FAUST human body meshes:

wget -O faust.tar.gz https://www.dropbox.com/s/jgm6hfif6evpi2b/faust.tar.gz?dl=0
tar -xvf faust.tar.gz

Then, preprocess the mesh segmentations:

./scripts/generate_data.sh

To overfit to a single mesh, run:

python scripts/train_reconstruction.py --data data/category --idx i --out out_dir

You should specify one of heads, torsos, arms, forearms, hands, or feet as category, and indicate an index between 0 and 99 as i to pick a mesh from the dataset.

To learn a minimal serfice, run:

python scripts/train_minimal.py --boundary boundary_config --idx i --out out_dir

Specify the boundary configuration boundary_config as either hopf, borromean, or trefoil.

To train a latent model, run:

python scripts/train_latent.py --data data/category --out out_dir

You should specify one of heads, torsos, arms, forearms, hands, or feet as category.

To monitor the training, launch a TensorBoard instance with --logdir out_dir.

Visualization

To render a turntable GIF from an overfit reconstruction or minimal surface model, run:

python scripts/render_current.py --infile out/model/it.pth --outfile out.gif

out/model/it.pth should be the checkpoint of a trained model.

To render a linear interpolation in boundary or latent space, run:

python scripts/render_interpolation.py --infile out/model/it.pth --outfile out.gif --data data/category --interpolation_type interpolation_type

out/model/it.pth should be the checkpoint of a trained model, and data/category the directory to the dataset used to train the model. You can choose between latent or boundary as the interpolation_type.

BibTeX

@article{palmer2021deepcurrents,
  title={{DeepCurrents}: Learning Implicit Representations of Shapes with Boundaries,
  author={Palmer, David and Smirnov, Dmitriy and Wang, Stephanie and Chern, Albert and Solomon, Justin},
  journal={arXiv:2111.09383},
  year={2021},
}