BEVPlace is a LiDAR-based place recognition method. It projects point clouds into Bird's-eye View (BEV) images and generate global feature with a group invariant network and the NetVLAD. Experiments show that BEVPlace significantly outperforms the state-of-the-art (SOTA) methods and generalizes well to previously unseen environments with little performance degradation. In addition, it can estimate postition of query point clouds by feature distance mapping. BEVPlace will certainly benefit various applications, including loop closure detection, global localization, and SLAM. Please feel free to use and enjoy it!
Create a conda environment and install pytorch according to you cuda version. Then install the dependencies by
pip install -r requirements.txt
The data of KITTI has been included in this repository. You can evaluate BEVPlace by simply running
python main.py
The recall rates will be displayed in the terminal.
Organize your own data following the description in data.md and custom you dataloader in dataset.py. Then evaluate the performance with the script main.py
Here are some experimental results on large-scale datasets.
- 2023-08-31: Update the pre-trained weights and the bev dataset of KITTI for reproducing the numbers in the paper.
- 2023-07-14: Our paper is accepted by ICCV 2023!
- 2023-03-14: Intial version
- 2022-09-02: Our method ranked 2nd in the General Place Recognition Competetion of ICRA 2022 (The 1st place solution is based on ensemble learning)!
@article{luo2023,
title={{BEVPlace}: {Learning LiDAR-based} Place Recognition using Bird's Eye View Images},
author={Lun, Luo and Shuhang, Zheng and Yixuan, Li and Yongzhi, Fan and Beinan, Yu and Siyuan, Cao and Hui-Liang, Shen},
journal={arXiv preprint arXiv:2302.14325},
year={2023}
}