This repository contains the official implementation for the paper:

Deformation Recovery: Localized Learning for Detail-Preserving Deformations Ramana Sundararaman, Nicolas Donati, Simone Melzi, Etienne Corman, Maks Ovsjanikov

1. Clone the repository:

   git clone https://github.com/geometry-processing/deformation-recovery.git
   cd deformation-recovery

   (Note: You might want to replace the URL with your actual repository URL)

2. Install dependencies: It is recommended to use a virtual environment (e.g., conda or venv).

   pip install -r requirements.txt

This project supports several datasets. To use them, you'll need to configure your data paths.

1. Download the datasets:

   • FAUST (website)
   • SURREAL (website)
   • SCAPE (website)
   • SHREC'19 (website)
   • DeformingThings4D (website)

2. Configure data paths: Copy the template file DATA_PATHS.py.template to DATA_PATHS.py:

   cp DATA_PATHS.py.template DATA_PATHS.py

   Then, edit DATA_PATHS.py to point to the locations of the datasets on your local machine.

To train the model, run the train.py script. Here's an example:

python train.py --exp_name faust_experiment --dataset faust_o --lr 1e-3 --n_epoch 120

Arguments:

• --exp_name: (required) Name for the experiment. Checkpoints will be saved in ./Logs/<exp_name>/.
• --dataset: The dataset to use for training. Choices: faust_r, faust_o, surreal, scape_r, scape_o, shrec19_r. (default: faust_o)
• --lr: Learning rate. (default: 1e-3)
• --n_epoch: Number of epochs. (default: 120)
• --chkpt: Path to a checkpoint to resume training from.
• --spec_inp: Use spectral input features.
• --pos_loss_weight: Weight for the position loss. (default: 10.0)
• --jac_loss_2_weight: Weight for the second Jacobian loss. (default: 0.5)

The --only_eval flag is available but not yet implemented. Evaluation functionality will be added soon.

If you find our work useful, please consider citing our paper:

@misc{sundararaman2024deformation,
      title={Deformation Recovery: Localized Learning for Detail-Preserving Deformations},
      author={Ramana Sundararaman and Nicolas Donati and Simone Melzi and Etienne Corman and Maks Ovsjanikov},
      year={2024},
      eprint={2410.08225},
      archivePrefix={arXiv},
      primaryClass={cs.GR}
}

This project is licensed under the MIT License.