RORNet: Partial-to-Partial Registration Network With Reliable Overlapping Representations

by Yue Wu , Yue Zhang , Wenping Ma , Maoguo Gong , Xiaolong Fan , Mingyang Zhang , A. K. Qin , and Qiguang Miao, and details are in paper.

Usage

1. Clone the repository.

2. Change the "DATA_DIR" parameter in the "data_utils.py" file to its own data set folder path.

3. Run the "main.py" in OverlapDetect file and save the pkl file; load pkl file trained by OverlapDetect file and run the OverlapReg file. Note: you need to make the "OverlapNet" model consistent for the OverlapDetect file and the OverlapReg file.

4. For convenience, We provide end-to-end training "running OverlapReg/main.py directly", but there may be a loss of accuracy.

Requirement

​ h5py=3.7.0

​ open3d=0.15.2

​ pytorch=1.11.0

​ scikit-learn=1.1.1

​ transforms3d=0.4.1

​ tensorboardX=1.15.0

​ tqdm

​ numpy

Dataset

​ (1) ModelNet40

​ (2) KITTI_odo

​ (3) Stanford Bunny

Citation

If you find the code or trained models useful, please consider citing:

@article{2023rornet,
  title={RORNet: Partial-to-Partial Registration Network With Reliable Overlapping Representations},
  author={Wu, Yue and Zhang, Yue and Ma, Wenping and Gong, Maoguo and Fan, Xiaolong and Zhang, Mingyang and Qin, AK and Miao, Qiguang},
  journal={IEEE Transactions on Neural Networks and Learning Systems},
  year={2023},
  publisher={IEEE}
}

Acknowledgement

Our code refers to PointNet, DCP and MaskNet. We want to thank the above open-source projects.