hmpr's Introduction

HMPR: A Hierarchical and Multi-modal Framework for Place Recognition with a Learnable Metric

Description

This is the source code for paper

C. Shu and Y. Luo. A Hierarchical and Multi-modal Framework for Place Recognition with a Learnable Metric, IEEE Transactions on Intelligent Vehicles.

Usage

build the ops

cd libs/pointops && python setup.py install && cd ../../

prepare the data
We use the processed point cloud submaps from PointNetVLAD.
The corresponding RGB images are retrieved from center camera of the original Oxford RobotCar dataset according to the closest timestamps.
Then RGB images are downsampled to 320*240 resolution.
generate training queries

cd generate_queries && python generate_training_tuples.py

training in first step

cd train && python train_coarse.py --work_path ../log/coarse_net --config_path ../config/config_coarse.yaml

generate top K candidate

cd generate_queries && python generate_topK_candidates.py

training in refinement step

cd train && python train_fine.py --work_path ../log/fine_net --config_path ../config/config_fine.yaml

Acknowledgements

Code is built based on AdaFusion, PointNetVLAD, PointNet++, PointWeb and CVTNet.

Citation

If you find our work useful in your research, please consider citing:

@ARTICLE{10508992,
author={Shu, Chengfu and Luo, Yutao},
journal={IEEE Transactions on Intelligent Vehicles},
title={A Hierarchical and Multi-modal Framework for Place Recognition with a Learnable Metric},
year={2024},
pages={1-10},
doi={10.1109/TIV.2024.3394213}}

Recommend Projects