Using Learnt Features in Indirect Visual SLAM. [project]

Related publication:

@inproceedings{hyhuang2020rdvo,
  title={Monocular Visual Odometry using Learned Repeatability and Description},
  author={Huaiyang Huang, Haoyang Ye, Yuxiang Sun and Ming Liu},
  booktitle={IEEE International Conference on Robotics and Automation (ICRA)},
  year={2020},
  organization={IEEE}
}

Demo videos:

We have tested this library in Ubuntu 18.04 with CUDA 9.2 and cuDNN 7.6. Prerequisites for installation:

1. ROS (ros-base is enough)

apt-get install ros-melodic-ros-base

2. miscs for installation:

apt-get install python-wstool python-catkin-tools 

3. OpenCV3

apt-get install libopencv-dev

4. CUDA tookit, cuDNN (change the version of libtorch in install.sh according to CUDA version)

5. Pangolin (optional, for visualization)

6. evo (optional, for evaluation)

pip install evo --upgrade --no-binary evo

Initialize a workspace:

mkdir -p /EXAMPLE/CATKIN/WORK_SPACE
cd /EXAMPLE/CATKIN/WORK_SPACE

mkdir src
catkin init
catkin config --extend /opt/ros/melodic
catkin config --cmake-args -DCMAKE_BUILD_TYPE=Release
catkin config --merge-devel

Clone the code:

cd src
git clone [email protected]:hyhuang1995/sp_orb_slam.git

Use the installation script:

cd sp_orb_slam
./install.sh

We provide examples on the New Tsukuba and the EuRoC MAV dataset. To run the demo on the New Tsukuba sequences:

1. Download the dataset

2. Replace the /PATH/TO/TSUKUBA/DATASET in tsukuba.launch with where the sequence is decompressed:

<param name="data_path" value="/PATH/TO/TSUKUBA/DATASET" />

3. Launch:

roslaunch tsukuba.launch seq:=lamps

To run the demo on the EuRoC MAV dataset:

1. Download the sequences (ASL Format)

2. Replace the /PATH/TO/EUROC/DATASET/ in euroc_mono.launch with where the sequence is decompressed:

<param name="data_path" value="/PATH/TO/EUROC/DATASET/$(arg seq)/mav0/" />

3. Launch:

roslaunch euroc_mono.launch seq:=MH_05_difficult

The output trajectories would be saved to orb_ros/expr.

If evo is installed, we provide scripts for evaluating the VO performances.

roscd orb_ros
./scripts/evaluate_tsukuba.sh

or

./scripts/evaluate_euroc.sh

and the results would be saved to orb_ros/expr.

Our implementation is built on top of ORB-SLAM2, see the license in the source files for more details. The authors would like to thank Raul et al. for their great work.