A Coarse-to-Fine Place Recognition Approach using Attention-guided Descriptors and Overlap Estimation (ICRA 2024)

This repository represents the official implementation of the ICRA 2024 paper:

CFPR is a coarse-to-fine framework for LiARD-based place recognition, which utilizes global descriptors to propose place candidates and employs overlap prediction to determine the final match. [Paper]

Instructions

This code has been tested on the following environment:

• Python: 3.7.16
• PyTorch: 1.12.1
• CUDA: 10.2

Pretrained models can be found here. Please download the pretrained models and place them in the models folder.

We use the poses file provided by SEMANTICKITTI.

Requirements

We recommend using the specific version of spconv to avoid compatibility issues:

• spconv-cu102=2.1.25

To set up your environment, you can use the following commands:

conda create -n onv python=3.7
conda activate onv
conda install pytorch==1.12.1 torchvision==0.13.1 torchaudio==0.12.1 cudatoolkit=10.2 -c pytorch
pip install spconv-cu102==2.1.25
pip install tqdm open3d==0.9.0 tensorboardX pyyaml scikit-learn matplotlib

Docker

We provide the Docker image fuchencan/onv for convenience.

To run the Docker container and activate the conda environment, use the following commands:

docker pull fuchencan/onv
docker run -it fuchencan/onv /bin/bash
conda activate onv

Extract Features

To extract BEV features, run the following command:

python tools/utils/gen_bev_features.py

Evaluation

To evaluate the model, run:

python evaluate/evaluate.py

Training

For training the backbone network and overlap estimation network, please refer to BEVNet. To train the global descriptor generation network, use the following command:

python train/train.py

The function evaluate_coarse is specifically for evaluating the coarse searching method using global descriptors.

Acknowledgements

We would like to thank the authors of the following repositories for their contributions, which have greatly aided our work:

• PointNetVlad
• PointNetVlad-Pytorch
• OverlapTransformer

Citation

If you find this repository helpful, please cite our work:

@inproceedings{fu2024coarse,
  title={A Coarse-to-Fine Place Recognition Approach using Attention-guided Descriptors and Overlap Estimation},
  author={Fu, Chencan and Li, Lin and Mei, Jianbiao and Ma, Yukai and Peng, Linpeng and Zhao, Xiangrui and Liu, Yong},
  booktitle={2024 IEEE International Conference on Robotics and Automation (ICRA)},
  pages={8493--8499},
  year={2024},
  organization={IEEE}
}

TODO

• Upload pretrained models
• Update the code to ICRA version
• Add KITTI-360 dataset
• ...