peterWon / GLoc3D

GlobalLoc3D

Implementation of our paper "Global Localization in Large-scale Point Clouds via Roll-pitch-yaw Invariant Place Recognition and Low-overlap Global Registration" (https://ieeexplore.ieee.org/document/10275135) in PyTorch.

Setup

Dependencies

1. PyTorch
2. Faiss
3. scipy
   • numpy
   • sklearn
   • h5py
4. tensorboardX

Usage

• Data preparation

  • Get into the folder of 'registration'

    • mkdir build && cd build && make -j8

  • Download KITTI Odometry and KITTI raw

  • Set 'root_dir' in 'kitti_i2i.py' to your own directory that saves KITTI raw

  • Set 'odometry_dir' in 'kitti_i2i.py' to your own directory that saves KITTI odometry

  • Generate index files using 'gen_index_files.py'

  • Run ./registration/build/save_probability_img YOUR_KITTI_RAW_DIR . You can find two folders saving processed images on the same directory of your root directory of KITTI raw.

• Training

  • Check arguments in main.py and run:
    • python main.py --mode=cluster --dataset=kitti --pooling=netvlad_fc
    • python main.py --mode=train --dataset=kitti --pooling=netvlad_fc

• Test

  • Place recognition
    • python main.py --mode=test --dataset=kitti --pooling=netvlad_fc --resume=YOUR_TRAINING_DIR --ckpt=best
  • Global localization
    • python main.py --mode=save_pt --dataset=kitti --pooling=netvlad_fc --resume=YOUR_TRAINING_DIR --ckpt=best
    • registration/build/global_localization VALSET_FILENAME GT_POSE_FILENAME MODEL_FILENAME

---

Acknowledgements

• The authors of Pytorch-NetVLAD
• The authors of Cartographer

license

Following licenses of the above acknowledged repositories.