radar_transformer

Transformer-based deep learning architecture for 3D point matching in sparse radar point clouds.

Cite (BibTeX)

If you use this software please cite:

@misc{michalczyk2025learningpointcorrespondencesradar,
      title={Learning Point Correspondences In Radar 3D Point Clouds For Radar-Inertial Odometry}, 
      author={Jan Michalczyk and Stephan Weiss and Jan Steinbrener},
      year={2025},
      eprint={2506.18580},
      archivePrefix={arXiv},
      primaryClass={cs.RO},
      url={https://arxiv.org/abs/2506.18580}, 
}

Remarks

If training on your data then make sure to include correct transformation between IMU and Radar sensors in the prepare_dataset.py script. This is because input to the network are pointclouds in the IMU frame. Of course, you can use any other common frame of reference. Also, make sure to adapt the training/inference script to your sensor's FOV and set the DC filtering offset in the utils.py script. DC offset is a constant detection close to (0, 0) caused by antennas cross-talk which sometimes occurs in consumer-grade FMCW radar sensors.