LC2: LiDAR-Camera Loop Constriants From Cross-Modal Place Recognition (RA-L 2023)

How to run

1. Preprocess the ViViD to contain imglist.txt and gpslist.txt

2. run manydepth by following command:

   python -m manydepth.test_all --target_image_path /your/vivid/path --intrinsics_json_path vivid_sequence_intrinsics.json --model_path manydepth/pretrained
   

   The script above will generate .npy and .png, representing depth in absolute scale and relative scale.

3. run lc2 by following command:

   python test_pr.py
   

Additional Funtionalities

• pr_curve.py: plots pr curve, given numpy files from test_pr.py.

• filter_and_visualze.py: gtsam-based pose optimizer using outputs from test_pr.py and 2D csv poses generated from rel. pose estimator.

** to run filter_and_visualize, you need the pose files below: loopposes.csv: Estimated rel.poses from EPnP, given depths from each depth values and local point cloud map. (2D form) odomposes.csv: Estimated poses from odometry, and projected into 2D.

Pretrained weights

1. Dual encoder [google drive]

2. Manydepth trained on ViViD [google drive]

###Codebases

[1] Patch-NetVLAD: Multi-Scale Fusion of Locally-Global Descriptors for Place Recognition (github)

[2] The Temporal Opportunist: Self-Supervised Multi-Frame Monocular Depth (github)

###Bibtex

@ARTICLE{lee2023lc2,
  author={Lee, Alex Junho and Song, Seungwon and Lim, Hyungtae and Lee, Woojoo and Myung, Hyun},
  journal={IEEE Robotics and Automation Letters}, 
  title={(LC)$^{2}$: LiDAR-Camera Loop Constraints for Cross-Modal Place Recognition}, 
  year={2023},
  volume={8},
  number={6},
  pages={3589-3596},
  doi={10.1109/LRA.2023.3268848}}