Skip to content

Latest commit

 

History

History
72 lines (44 loc) · 2.79 KB

README.md

File metadata and controls

72 lines (44 loc) · 2.79 KB
Raw

IFMatch

PyTorch implementation for our ECCV 2022 paper Implicit field supervision for robust non-rigid shape matching

Setup

  1. Setup the conda environment using from the ifmatch_env.yml as conda env create -f ifmatch_env.yml and activate it.

  2. Install the pytorch-meta as cd pytorch-meta && python setup.py install

Dataset and Pre-Processing

  1. We provide the datasets and variants used in our paper here

  2. Once the dataset have been downloaded, we have two-staged pre-processing,

    1. Sampling SDF: To sample points with SDF, we follow the DeepSDF scheme as given here. Place all the npz files into (say) /path/to/npz.

    2. Sampling surface with normal: For this, we use the mesh-to-sdf package. To perform this step, run data_process.py by providing the path to ply files and the npz files from previous point. Run with --help option to know other required parameters.

  3. Once the pre-processing is done, your data directory should have three directories, free_space_pts containing the SDF, surface_pts_n_normal containing the surface points along with normal information and vertex_pts_n_normal containing vertex points.

  4. Step 2 is repeated for both training and test dataset alike.

  5. We provide pre-processed samples consisting of test-set shapes from the FAUST-Remesh dataset here: https://nuage.lix.polytechnique.fr/index.php/s/gb8D3KHBeb7zqNL.

Training

To train, run the following by appropriately replacing parameters,

python train.py --config configs/train/<dataset>.yml --split split/train/dataset.txt --exp_name <my_exp>

Evaluation

Our evaluation is two staged, first we find the optimal latent vector (MAP), then we solve for the P2P map between shapes

  1. To run the MAP step,
python evaluate.py --config configs/eval/<dataset.yml>
  1. To obtain the P2P map,
python run_matching.py --config configs/eval/<dataset.yml> --latent_opt

Pre-trained Models

We perform 3 distinct training in total for reported results in the paper. Respective models can be downloaded from here

Citation

If you find our work useful, please cite the arxiv version below. (To be updated soon...)

@misc{sundararaman2022implicit,
    title={Implicit field supervision for robust non-rigid shape matching},
    author={Ramana Sundararaman and Gautam Pai and Maks Ovsjanikov},
    year={2022},
    eprint={2203.07694},
    archivePrefix={arXiv},
    primaryClass={cs.CV}
}

Acknowledgements

We thank authors of DIF-Net and SIREN for graciously open-sourcing their code.