HPPLO-Net: Unsupervised LiDAR Odometry Using a Hierarchical Point-to-Plane Solver

This is the official implementation of IEEE Transactions on Intelligent Vehicles 2023 paper "HPPLO-Net: Unsupervised LiDAR Odometry Using a Hierarchical Point-to-Plane Solver" created by Beibei Zhou, Yiming Tu, Zhong Jin, Chengzhong Xu, Hui Kong.

Citation

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

  @ARTICLE{10160144,
    author={Zhou, Beibei and Tu, Yiming and Jin, Zhong and Xu, Chengzhong and Kong, Hui},
    journal={IEEE Transactions on Intelligent Vehicles}, 
    title={HPPLO-Net: Unsupervised LiDAR Odometry Using a Hierarchical Point-to-Plane Solver}, 
    year={2023},
    volume={},
    number={},
    pages={1-13},
    doi={10.1109/TIV.2023.3288943}}

Prequisites

Our model is trained and tested under:

• Python 3.7.0
• NVIDIA GPU + CUDA CuDNN
• PyTorch (torch >= 1.2.0)
• scipy
• tqdm
• sklearn
• numba
• cffi
• pypng
• pptk

Usage

Datasets

We use KITTI odometry dataset in our experiments.

Data preprocessing

• remove the ground points of pointclouds by running groundtest.m located in the directory data_preprocess_zbb/matlab_ground/devkit/matlab using MATLAB.
• downsample the pointclouds by running the file zbb_data_process.py located in the directory data_preprocess_zbb/python_downsample using python.

python zbb_data_process.py

Pay attention to modifying the file paths.

Training

Train the network by running

python traincomer.py

Please reminder to specify the onlyneedtest(False), loadmodel(False),dataroot,trainset(sequences for training), batch_size in param/titanrtx_1.py.

Testing

Test the network by running

python traincomer.py

Please reminder to specify the onlyneedtest(True), loadmodel(True),model(path to HPPLO-Net model), dataroot,testset(sequences for testing), testbatch in param/titanrtx_1.py.

Acknowledgments

We thank the following open-source projects for the help of the implementations:

• PointNet++
• KITTI_odometry_evaluation_tool
• PointPWC
• FlowNet3D