Project page: https://sites.google.com/view/dsl-ram-lab/

Related publications:

@inproceedings{ye2020monocular,
  title={Monocular direct sparse localization in a prior 3d surfel map},
  author={Ye, Haoyang and Huang, Huaiyang and Liu, Ming},
  booktitle={2020 IEEE International Conference on Robotics and Automation (ICRA)},
  pages={8892--8898},
  year={2020},
  organization={IEEE}
}
@inproceedings{ye20213d,
  title={3D Surfel Map-Aided Visual Relocalization with Learned Descriptors},
  author={Ye, Haoyang and Huang, Huaiyang and Hutter, Marco and Sandy, Timothy and Liu, Ming},
  booktitle={2021 International Conference on Robotics and Automation (ICRA)},
  pages={5574-5581},
  year={2021},
  organization={IEEE}
}

Video: https://www.youtube.com/watch?v=LTihCBGcURo

1. Pangolin.
2. CUDA.
3. Ceres-solver.
4. PCL, the default version accompanying by ROS.
5. OpenCV, the default version accompanying by ROS.

0. git submodule update --init --recursive
1. mkdir build && cd build
2. cmake .. -DCMAKE_BUILD_TYPE=RelWithDebInfo
3. make -j8

The sample config file can be downloaded from this link.

To run the example:

[path_to_build]/src/dsl_main --path "[path_to_dataset]/left_pinhole"

1. Collect LiDAR and camera data.
2. Build LiDAR map and obtain LiDAR poses (the poses are not necessary).
3. Pre-process LiDAR map to make the [path_to_dataset]/*.pcd map file contains normal_x, normal_y, normal_z fields (downsample & normal estimation).
4. Extract and undistort images into [path_to_dataset]/images.
5. Set the first camera pose to initial_pose and other camera parameters in [path_to_dataset]/config.yaml.

This implementation of DSL takes Ceres Solver as backend, which is different from the the implementation of the original paper with DSO-backend. This leads to different performance, i.e., speed and accuracy, compared to the reported results.

This work is inspired from several open-source projects, such as DSO, DSM, Elastic-Fusion, SuperPoint, DBoW2, NetVlad, LIO-mapping and etc.

The source code is released under GPL-3.0.