Partial Wasserstein Adversarial Network for Non-rigid Point Set Registration

---

Update 2023/03/22: In the near future, several updates will be made:

1. A new version for rigid point set registration will be released.
2. API will be polished so that this code can be used more easily.

---

This is a Pytorch implement of the PWAN model proposed in Partial Wasserstein Adversarial Network for Non-rigid Point Set Registration.

Partial Wasserstein Adversarial Network (PWAN) seeks to partially match two unnormalized distributions. It can be seen as a generalization of the popular Wasserstein GAN model to the unbalanced case. The code in this repository applies PWAN to point set registration problems.

Point sets	Potential	Gradient

Requirements

• pytorch=1.4
• open3d (for visualizing 3D point sets)
• matplotlib (for visualizing 2D point sets)
• pykeops (for fine-tuning registration)

Usage

Common usage is shown in scrip.py, and see more detailed explanations of the parameters in main.py. Note the algorithm may sometimes fail to converge due to the regularizer term, try using smaller regularizer (e.g. larger sigma or smaller lambda) to stabilize the training.

Reference

@inproceedings{wang2022partial,
    title={Partial Wasserstein Adversarial Network for Non-rigid Point Set Registration},
    author={Zi-Ming Wang and Nan Xue and Ling Lei and Gui-Song Xia},
    booktitle={International Conference on Learning Representations (ICLR)},
    year={2022}
}

For any question, please contact me ([email protected]).