Correspondence Learning via Linearly-invariant Embedding (PyTorch)

This repository is a PyTorch implementation of Correspondence Learning via Linearly-invariant Embedding.

This is not the code used to produce the paper results, which can be found Here. This implementation has been made to make handier the use of the method (and also to replicate it).

Requirements

To install requirements:

pip install -r requirements.txt

Installing PyTorch may require an ad hoc procedure, depending on your computer settings.

Data & Pretrained models

You can download data and the pre-trained models using the scripts:

python .\data\download_data.py
python .\models\pretrained\download_pretrained.py

Training

To train the basis and descriptors models, run these commands:

python .\code\train_basis.py
python .\code\train_desc.py

Evaluation

To evaluate the model on FAUST w\noise, run:

python .\code\test_faust.py

And in matlab the script:

.\evaluation\evaluation.m

Results

These are the results of the two implementations:

Model name	Ours	Ours+Opt
TF 1.5	6.0e-2	2.9e-2
PyTorch	5.7e-2	3.1e-2

The small discrepancies have several reasons:

1. basis and descriptors networks are trained 400 epochs in PyTorch implementation; several thousand in TF 1.5
2. while the two implementations are similar, there are some differences in the training process and hyperparameters due to libraries.
3. the training requires pseudo-inverse computation; these can produce different results depending on the library

License

If you use this code, please cite our paper.

@article{marin2020correspondence,
  title={Correspondence learning via linearly-invariant embedding},
  author={Marin, Riccardo and Rakotosaona, Marie-Julie and Melzi, Simone and Ovsjanikov, Maks},
  journal={Advances in Neural Information Processing Systems},
  volume={33},
  year={2020}
}

This work is licensed under a Creative Commons Attribution-NonCommercial 4.0 International License. For any commercial uses or derivatives, please contact us.