C2L-PR

The official implementation of C2L-PR: Cross-modal Camera-to-LiDAR Place Recognition via Modality Alignment and Orientation Voting.

This work has been accepted by IEEE T-IV 2024. tada: Huaiyuan Xu; Huaping Liu; Shoudong Huang; Yuxiang Sun

Overview

The figure illustrates our idea of cross-modal place recognition with a visual image against a pre-built LiDAR point-cloud database. Given an on-line query image captured by a vehicle-mounted monocular camera, a semantic point cloud is generated from the image to alleviate the modality gap to find the matched reference point cloud in the database.

Demo video (best viewed in 1080p quality): https://www.youtube.com/watch?v=60S9BFzgWI0

Get Started

Installation and Data Preparation

step 1. Please prepare environment as that in Docker.

$ cd docker
$ docker build -t your_docker_image .
$ docker run -it --shm-size 8G -p 1001:6006 -p 1000:22 --ipc host --name your_docker_container --gpus all -v ~/your_project_direction:/workspace your_docker_image  /bin/bash 

step 2. Prepare C2L-PR repo by.

$ git clone https://github.com/lab-sun/C2L-PR.git
$ cd C2L-PR

step 3. Download data and checkpoints. Then arrange the folder as:

C2L-PR/
├── checkpoints
├── data
    └── kitti 
        ├── 0
        ├── 1
        ├── 2
        ├── 3
        ├── 4
        ├── 5
        ├── 6
        ├── 7
        ├── desc_image
        ├── gt_split90
        ├── pairs_kitti
        └── pose_kitti
├── docker
└── images
    ...

Train models

In train.py, please modify os.environ['CUDA_VISIBLE_DEVICES'], pointing to a specifc graphic card. Please modify directions of 'gt_folder','velo_desc_folder_0' to '7', and 'img_desc_folder', appropriately. Then, run the following command:

$ sh train.sh

Evaluate models

In eval_f1.py, please modify os.environ['CUDA_VISIBLE_DEVICES']. In eval_f1.sh, please modify '--model', '--velo_desc_file_0' to '7', '--img_desc_file', '--pairs_file', appropriately.

# Evaluate place recognition. We take F1 score as the metric.
$ sh eval_f1.sh
# Evaluate loop closure. We take recall@K as the metric.
$ sh eval_recall.sh

Acknowledgement

We thank the fantastic work RINet for its pioneer code release, which provide codebase for C2L-PR.

Citation

If you use C2L-PR in an academic work, please cite our paper:

@ARTICLE{10586273,
author={Xu, Huaiyuan and Liu, Huaping and Huang, Shoudong and Sun, Yuxiang},
journal={IEEE Transactions on Intelligent Vehicles}, 
title={C2L-PR: Cross-modal Camera-to-LiDAR Place Recognition via Modality Alignment and Orientation Voting}, 
year={2024},
volume={},
number={},
pages={1-17},
doi={10.1109/TIV.2024.3423392}}

Contact

huaiyuan.xu@polyu.edu.hk

Website: https://www.labsun.org/