AirDOS: Dynamic SLAM benefits from Articulated Objects

TartanAir Shibuya Dataset

You may download the dataset from: https://github.com/haleqiu/tartanair-shibuya

Instruction

Dependencies

• Eigen (Tested on v3.3)
• OpenCV (Tested on Opencv 3.4)
• Pangolin (v0.5)

Build

Run the ./build.sh in root directory to build third party libraries, AirDOS and construct a binary representation of ORB vocabulary file.

Run an example script

Currently, the project only supports Stereo mode. The stereo_human executable in Examples/Stereo

./Examples/Stereo/stereo_human \
  ./Vocabulary/ORBvoc.txt \
  ./Examples/Stereo/config/tartanair.yaml \
  /.../TartanAir_shibuya/RoadCrossing07 \
  ./Evaluation/data/trajectory_output.txt

Evaluate Trajectory

For evaluation, please check https://github.com/castacks/tartanair_tools.git.

Note: the TartanAir Tools expect to read the trajectory format with 7 columns, representing the translation and quaternion.

The trajectory exported by the program contains 8 columns, where first column is the timestamp of pose.

To evaluate the exported trajectory using TartanAir Tools, you need to remove the first column of exported trajectory file.

We also provided an evaluation tool based on evo package, for more detail, please check this README.md.

Publications

AirDOS: Dynamic SLAM benefits from Articulated Objects

@inproceedings{qiu2022airdos,
  author={Qiu, Yuheng and Wang, Chen and Wang, Wenshan and Henein, Mina and Scherer, Sebastian},
  booktitle={IEEE International Conference on Robotics and Automation (ICRA)}, 
  title={AirDOS: Dynamic SLAM benefits from Articulated Objects}, 
  year={2022},
  pages={8047-8053},
  doi={10.1109/ICRA46639.2022.9811667}
 }

Contributors

haleqiu, MarkChenYutian