To get started:
git clone [email protected]:arpg/bo_icp.git
Install the package:
cd bo_icp
pip install .
Note it's recommended to do this in virtual environment such as conda.
Experiments from the papers were performed on the KITTI and TUM Datasets.
For the KITTI dataset use format: 'KITTI' in the config file and for TUM use format: 'tum'. For all other file types the format tag gets passed into the open3d.io.read_point_cloud() function. The auto specifier will work for PCD files. For other file types please refer to the library documentation for more information.
The BO_ICP_example.ipynb shows an example of how to use this code. Point cloud locations can either be specified in the config file or by passing in a source_cloud and target_cloud path to the BO_ICP constructor (as done in the notebook).
Configs from the paper are provided as config_a, config_b, and config_c. It's recommended to start with b for a new dataset. If you need more accuracy try C and for a faster runtime try A. The generic config.yaml has comments describing each of the parameters.
We make use of the BayesianOptimization library by Fernando Nogueira and Open3d.
If you find this work useful please consider citing:
@INPROCEEDINGS{10160570,
author={Biggie, Harel and Beathard, Andrew and Heckman, Christoffer},
booktitle={2023 IEEE International Conference on Robotics and Automation (ICRA)},
title={BO-ICP: Initialization of Iterative Closest Point Based on Bayesian Optimization},
year={2023},
volume={},
number={},
pages={6944-6950},
doi={10.1109/ICRA48891.2023.10160570}}