Implementation of IROS20 paper - "Semantic Graph Based Place Recognition for 3D Point Clouds"
Code for IROS2020 paper Semantic Graph Based Place Recognition for 3D Point Clouds
Pipeline overview.
If you think this work is useful for your research, please consider citing:
@inproceedings{kong2020semantic,
title={Semantic Graph based Place Recognition for Point Clouds},
author={Kong, Xin and Yang, Xuemeng and Zhai, Guangyao and Zhao, Xiangrui and Zeng, Xianfang and Wang, Mengmeng and Liu, Yong and Li, Wanlong and Wen, Feng},
booktitle={2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)},
pages={8216--8223},
year={2020},
organization={IEEE}
}
We recommend python3.6. You can install required dependencies by:
pip install -r requirements.txt
The data structure is:
data
|---00
| |---000000.json
| |---000001.json
| |---...
|
|---01
| |---000000.json
| |---000001.json
| |---...
|
|---...
|
|---00.txt
|---01.txt
|---...
You can download the provided preprocessed data. Or you can refer to the 'data_process' dir for details of generating graphs.
Before training the model, you need to modify the configuration file in ./config according to your needs. The main parameters are as follows:
After preparing the data and modifying the configuration file, you can start training. Just run:
python main_sg.py
This example takes a pair of graphs as input and output their similarity score. To run this example, you need to set the following parameters:
Then just run:
python eval_pair.py
This example tests a sequence, the results are it's PR curve and F1 max score. To run this example, you need to set the following parameters:
Then just run:
python eval_batch.py
We provide the raw data of the tables and curves in the paper, including compared methods M2DP, PointNetVLAD, Scan Context.
We recommend users refer the work SSC for a fair comparison with recent methods in the same data distribution.
Please refer to our modified repo for training and testing PointNetVLAD on KITTI dataset, which is mentioned in our paper as PV_KITTI.
Thanks to the source code of some great works such as SIMGNN, DGCNN.