Feature-metric registration

This repository is the implementation of our CVPR 2020 work: "Feature-metric Registration: A Fast Semi-supervised Approach for Robust Point Cloud Registration without Correspondences"

<p align="center"> <img width="50%" src="https://github.com/XiaoshuiHuang/xiaoshuihuang.github.io/blob/master/research/2020-feature-metric.png?raw=true" /></p> <p align="center"><img width="50%" src="https://github.com/XiaoshuiHuang/xiaoshuihuang.github.io/blob/master/research/2020-feature.png?raw=true" /></p>

There are several lights of this work:

1. 💡 This work solves the point cloud registration using feature-metric projection error.

2. 💡 This work can be trained with unsupervised or semi-supervised manner.

3. 💡 This work can handle both high noise and density variations.

4. 💡 This work is potential to handle cross-source point cloud registration.

To run the code, please follow the below steps:

1. Install dependencies:

pip install torch===1.5.1 torchvision===0.6.1 -f https://download.pytorch.org/whl/torch_stable.html argparse numpy glob matplotlib six 

2. Train the model

2.1. Train on dataset ModelNet40:

python train.py -data modelnet

2.2. Train on dataset 7scene:

python train.py -data 7scene

3. Evalute the model

3.1. Evaluate on dataset ModelNet40:

python evalute.py -data modelnet

3.2. Evaluate on dataset 7scene:

python evalute.py -data 7scene

4. Pre-trained models

The pretrained models are stored in the result folder.

5. Code for testing your own point clouds

Test your own point clouds by running: 

python demo.py

You need to change the path0 and path1 of demo.py to the paths of your own  point clouds.

6. Citation

@InProceedings{Huang_2020_CVPR,
    author = {Huang, Xiaoshui and Mei, Guofeng and Zhang, Jian},
    title = {Feature-Metric Registration: A Fast Semi-Supervised Approach for Robust Point Cloud Registration Without Correspondences},
    booktitle = {The IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    month = {June},
    year = {2020}
}

Acknowledgement

We would like to thank the open-source code of AtlasNet and pointnetlk