DeepMapping2: Self-Supervised Large-Scale LiDAR Map Optimization
Chao Chen*, Xinhao Liu*, Yiming Li, Li Ding, Chen Feng
News
[2023-06] π₯ Quaternion has been added as the default rotation representation, resulting in faster convergence from 30 epochs to 7 epochs.
[2023-06] π₯ Detailed instruction is available for implementing DeepMapping2 on a custom dataset.
[2023-03] The camera-ready version of our paper is available at arXiv. New figures are added to the supplementary material.
[2023-02] Our paper is accepted by CVPR 2023.
[2022-12] Our paper is available at arXiv and the project page is online.
Abstract
LiDAR mapping is important yet challenging in self-driving and mobile robotics. To tackle such a global point cloud registration problem, DeepMapping converts the complex map estimation into a self-supervised training of simple deep networks. Despite its broad convergence range on small datasets, DeepMapping still cannot produce satisfactory results on large-scale datasets with thousands of frames. This is due to the lack of loop closures and exact cross-frame point correspondences, and the slow convergence of its global localization network. We propose DeepMapping2 by adding two novel techniques to address these issues: (1) organization of training batch based on map topology from loop closing, and (2) self-supervised local-to-global point consistency loss leveraging pairwise registration. Our experiments and ablation studies on public datasets (KITTI, NCLT, and Nebula) demonstrate the effectiveness of our method. Our code will be released.
Getting Started:
Installation
The code is tested with Python 3.9, PyTorch 1.13.1, and CUDA 11.6.
To install the dependencies, you can create a virtual environment with
conda env create -f environment.yml
Data Preparation
To download the dataset used for training and testing, please refer to ./data/README.md
Usage
To train the model, execute the script
cd script/
./run_train.sh
The visualization and evaluation results will be saved in the results folder.
To use a different initial pose and pairwise registration, please edit INIT and PAIRWISE in the script to direct to the corresponding files.
Citation
If you find this work useful for your research, please cite our paper:
@inproceedings{chen2023deepmapping2,
title={DeepMapping2: Self-Supervised Large-Scale LiDAR Map Optimization},
author={Chen, Chao and Liu, Xinhao and Li, Yiming and Ding, Li and Feng, Chen},
booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
pages={9306--9316},
year={2023}
}