HypLiLoc

(CVPR 2023) HypLiLoc: Towards Effective LiDAR Pose Regression with Hyperbolic Fusion

https://arxiv.org/abs/2304.00932

proceddings

---

You can also view at https://youtu.be/qplZMOZG-7k

💥💥:racehorse::racehorse: We have refined the code structure. This new version can run at 80FPS on NVIDIA 3090 GPU or 150FPS on NVIDIA 4090 GPU!

• Requirements

  PyTorch installation (You may also use Pytorch2.0, which is also compatible):

  pip install torch==1.11.0+cu113 torchvision==0.12.0+cu113 torchaudio==0.11.0 --extra-index-url https://download.pytorch.org/whl/cu113
  

  You can also use the following script:

  conda install pytorch==1.11.0 torchvision==0.12.0 torchaudio==0.11.0 cudatoolkit=11.3 -c pytorch
  

  Other dependencies:

  colour_demosaicing==0.2.2
  geotorch==0.3.0
  matplotlib==3.5.3
  numpy==1.19.5
  open3d==0.15.2
  opencv_python==4.6.0.66
  Pillow==9.3.0
  scipy==1.9.1
  setuptools==63.4.1
  tqdm==4.64.0
  transforms3d==0.4.1
  

• Extra dependency pointnet2 installation

  cd network/pointnet2
  python setup.py install
  cd ..
  cd ..
  

• Platform

  Ubuntu 20.04
  CUDA 11.6/11.8
  python 3.8
  

• Dataset

  We currently provide the Oxford Radar dataset that has been pre-processed.

  google drive

  After downloading, you can unzip it and record the path, e.g.

  /home/workstation/Radar_RobotCar/
  

  Tips: You can better put the dataset under some folder supported by SSD to achieve fast reading speed.

• Trained weights

  We provide the trained optimal weights for the Full-8 route.

  https://drive.google.com/file/d/1xunKg82BK2-yOyh7q04AOxL6qL5VKEQE/view?usp=share_link

  After downloading, you can unzip it and put it under this repo's root, which will be like:

  HypLiLoc/logs
  

• Inference on the Full-8 route

  We have refined the code structure, and the version can run at 140FPS !

  python eval.py --data_dir /home/workstation/Radar_RobotCar/ --cuda 0 --scene full8 
  

• Train

  python train.py --data_dir /home/workstation/Radar_RobotCar/ --cuda 0 --scene full8 
  

• Other information:

  You can view tools/options.py to set running arguments.

• Citation

  @inproceedings{wang2023hypliloc,
    title={HypLiLoc: Towards Effective LiDAR Pose Regression with Hyperbolic Fusion},
    author={Wang, Sijie and Kang, Qiyu and She, Rui and Wang, Wei and Zhao, Kai and Song, Yang and Tay, Wee Peng},
    booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
    pages={5176--5185},
    year={2023}
  }
  

• Code reference

  https://github.com/sijieaaa/RobustLoc

  https://github.com/htdt/hyp_metric

  https://github.com/BingCS/AtLoc