README.md

October 26, 2025 · View on GitHub

Deep Visual Odometry for Stereo Event Cameras

Video

[1] Deep Visual Odometry for Stereo Event Cameras, Sheng Zhong, Junkai Niu, Yi Zhou, IEEE Robotics and Automation Letters (RA-L), 2025. PDF, Video.

1. Installation

We have tested SDEVO on machines with the following configurations

Ubuntu 20.04 LTS + ROS Noetic + OpenCV 4.2 + Eigen 3.3.9 + CUDA Toolkit 11.x

1.1 C++ node

First, create a catkin workspace and and download the code into the src directory.

cd ~/catkin_ws/src
git clone https://github.com/NAIL-HNU/SDEVO.git

Then, configure the environment and install dependencies by following the ESVO2 repository. Specific dependency items can be found in the dependencies.yaml.

Finally, compile it.

cd ~/catkin_ws
catkin_make

1.2 Python node

We refer to DEVO repository and use Anaconda to manage the Python environment.

First, create and activate the Anaconda environment

conda env create -f environment.yml
conda activate sdevo

Then, install the package

cd ~/DEVO
pip install .

2. Usage

Since the rosbag with event data is not provided in the datasets, we repackage the required data as the input for the system. You can access most of the rosbag files we repacked through the

After you get the repackaged data, you can try running it using the following command.

cd ~\catkin_ws
source devel/setup.bash
conda activate sdevo
roslaunch image_representation voxel_xxx.launch

This will launch two image_representation nodes (for left and right event cameras, respectively), the sdevo node simultaneously. Then play the input (already downloaded) bag file by running

rosbag play xxx.bag --clock

The trajectories will be saved in the path in /output/poses_interpolated.txt.

If your hardware cannot support real-time execution of our system, you may modify the rosbag playback rate. However, you must correspondingly adjust the trigger frequency of the periodic signal in the launch file to maintain synchronization:

<node name="global_timer" pkg="rostopic" type="rostopic" args="pub -s -r 2 /sync std_msgs/Time 'now' ">

The frequency of the global_timer (in Hz) divided by the rosbag playback rate must equal the voxel generation frequency (generation_rate_hz). This relationship maintains temporal consistency between data playback and system processing.

Implementation Notes:

When decreasing playback rate (slower than real-time), proportionally decrease the global_timer frequency

When increasing playback rate (faster than real-time), proportionally increase the global_timer frequency

Always verify the resulting voxel generation rate matches system requirements

3. Abstract

We present a deep learning-based method for visual odometry using stereo event cameras. The proposed system is built on top of DEVO, a monocular event-only VO system that leverages a learned patch selector and a pooled multinomial sampling for tracking sparse event patches.