Awesome

The official implementation of the CVPR'2023 paper ([arXiv]).

Setting up the python interpreter

• We recommend installing the required python packages directly via anaconda:

conda env create --name gmsm_env -f gmsm_env.yml
conda activate gmsm_env

Data

• The repository contains sample code to train our full model on SHREC'20. Please download the high-resolution data from here and preprocess it, following the steps from the DeepShells repository.
• Save the processed shapes under data/shrec20, or specify the path in utils/data.py.
• Other datasets can be trained analogously, following the same preprocessing steps. In our experiments, we considered the nearly-isometric datasets FAUST remeshed, SCAPE remeshed, SURREAL, SHREC'19, the topological datasets SMAL, TOPKIDS, SHRECISO as well as the non-isometric datasets SHREC'20, SMAL, TOSCA. For each dataset, running the script preprocess_data/preprocess_dataset.m performs all the necessary preprocessing steps.

Run the code

Train

• To train our model, simply run

python3 main.py

• The outputs will be saved automatically in a new folder under results/shrec20.

Test

• To test our model, run the evaluation script

python3 eval_scripts.py

• Results will be saved under results/shrec20_pretrained.
• The script outputs the query correspondences for all pairs in individual files, as well as the predicted shape graph and the validation losses for all pairs.
• By default, a pretrained version of our model is used to produce the query correspondences on SHREC'20.
• To evaluate different runs, specify the configuration in eval_scripts.py.

Citation

If you use our implementation for your own work, please cite:

@article{eisenberger2023g,
  title={G-MSM: Unsupervised Multi-Shape Matching with Graph-based Affinity Priors},
  author={Eisenberger, Marvin and Toker, Aysim and Leal-Taix{\'e}, Laura and Cremers, Daniel},
  journal={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  year={2023}
}