Tutorial: Run OpenVINS on the EuRoC MAV Dataset

This tutorial walks you through running OpenVINS end-to-end on the EuRoC MAV Dataset, the most widely used benchmark for visual-inertial odometry. EuRoC provides synchronized monochrome stereo images at 20 Hz from two global-shutter cameras alongside a MEMS ADIS16448 IMU at 200 Hz. OpenVINS ships a ready-made configuration for this dataset, so no manual calibration is required. Before starting, make sure you have completed the installation guide and have a working build.

Configuration files

All OpenVINS parameters are driven by three YAML files. For EuRoC these are already provided under config/euroc_mav/ in the repository:

File	Purpose
`estimator_config.yaml`	Main estimator settings: feature counts, SLAM options, zero-velocity update thresholds, etc.
`kalibr_imu_chain.yaml`	IMU noise densities, random walk parameters, and ROS topic names
`kalibr_imucam_chain.yaml`	Camera-to-IMU extrinsics, time offset, and camera intrinsics

You can override any estimator_config.yaml parameter directly from the launch file without editing the YAML file on disk.

Launch files

The ov_msckf/launch/ directory contains the following files:

File	Description
`subscribe.launch`	ROS 1 launch — subscribes to live or bagged camera and IMU topics
`subscribe.launch.py`	ROS 2 launch — equivalent subscriber node
`simulation.launch`	Runs the built-in simulator (no dataset required)
`serial.launch`	Processes images in serial (useful inside Docker or slow machines)
`display.rviz`	RViz configuration for ROS 1
`display_ros2.rviz`	RViz configuration for ROS 2
`display_driving.rviz`	RViz configuration for driving/vehicle datasets

Download the EuRoC bag

Download the V1_01_easy sequence (Vicon Room 1, easy trajectory) — a good starting point because it has low-dynamic motion and good lighting.

ROS 1 bag
ROS 2 bag

Download directly from the ASL server:

wget http://robotics.ethz.ch/~asl-datasets/ijrr_euroc_mav_dataset/vicon_room1/V1_01_easy/V1_01_easy.bag

The ROS 2 version is a converted bag in the db3 + YAML format hosted on Google Drive. Download and unzip it:

# Download from:
# https://drive.google.com/file/d/1LFrdiMU6UBjtFfXPHzjJ4L7iDIXcdhvh/view
# Then unzip to get the V1_01_easy/ folder containing the .db3 and metadata.yaml files.

A ROS 2 bag is a directory (not a single file) containing a .db3 SQLite database and a metadata.yaml file. Pass the folder name — not a file path — to ros2 bag play.

Launch OpenVINS

ROS 1
ROS 2

Open three separate terminals. In the first, start a persistent ROS master so you do not have to restart RViz on each run:

# Terminal 0: ROS master
roscore

In a second terminal, source the workspace and launch the estimator:

# Terminal 1: OpenVINS estimator
source devel/setup.bash
roslaunch ov_msckf subscribe.launch config:=euroc_mav

The config:=euroc_mav argument resolves to config/euroc_mav/estimator_config.yaml automatically. The node name is run_subscribe_msckf and it subscribes to the camera and IMU topics listed in the config.You can override individual parameters at launch time without editing any file:

roslaunch ov_msckf subscribe.launch config:=euroc_mav use_stereo:=true max_cameras:=2

In the remaining two terminals, start RViz and then play the bag:

# Terminal 2: visualization
rviz -d ov_msckf/launch/display.rviz

# Terminal 3: data playback
rosbag play V1_01_easy.bag

Source the install space and launch the estimator:

# Terminal 1: OpenVINS estimator
source install/setup.bash
ros2 launch ov_msckf subscribe.launch.py config:=euroc_mav

The ROS 2 launch file dynamically locates the config directory via ament_index_python, so the config:=euroc_mav argument works the same way as in ROS 1. Valid configs are any folder name present in the installed config/ directory.In a second terminal, play the converted bag:

# Terminal 2: data playback
ros2 bag play V1_01_easy

Visualize in RViz

Once the bag starts playing you should see the following topics published in RViz:

Feature tracks — colored point cloud of tracked 2D feature positions projected into 3D
SLAM landmarks — persistent 3D map points maintained in the filter state
Pose estimate — 6-DOF camera/IMU pose as a TF frame and a PoseWithCovarianceStamped
Path — history of pose estimates drawn as a line

Open ov_msckf/launch/display.rviz (or display_ros2.rviz for ROS 2) from within RViz using File → Open Config if it does not load automatically.

Configuration overview

The estimator_config.yaml file exposes all key tuning knobs. Below are the most commonly adjusted parameters:

Parameter	Default	Description
`use_stereo`	`true`	Use both cameras in a stereo configuration
`max_cameras`	`2`	Number of cameras to use
`num_pts`	`200`	Maximum number of MSCKF feature tracks per frame
`num_slam`	`50`	Number of persistent SLAM features to maintain
`dt_slam_delay`	`2.0`	Seconds after init before SLAM features are added
`try_zupt`	`false`	Enable zero-velocity update
`zupt_max_velocity`	`1.0`	Max velocity (m/s) below which ZUPT is attempted
`gravity_mag`	`9.81`	Local gravity magnitude in m/s²
`up_msckf_sigma_px`	`1.0`	Pixel noise standard deviation for MSCKF updates
`up_slam_sigma_px`	`1.5`	Pixel noise standard deviation for SLAM updates

Start with the default EuRoC configuration before changing any parameters. The provided values are well-tuned for this dataset. When tuning for a new sensor, focus first on num_pts, up_msckf_sigma_px, and IMU noise parameters in kalibr_imu_chain.yaml.

Tuning tips

Estimator diverges or drifts quickly

Check that the IMU noise parameters in kalibr_imu_chain.yaml match your sensor’s datasheet. Overly optimistic (too small) noise values cause the filter to over-trust the IMU and reject valid visual updates. Run kalibr_imu_validator or imu_utils to measure noise densities on your hardware.

Feature tracking drops to zero

This usually means the camera topic names in kalibr_imucam_chain.yaml do not match the topics in your bag. Run rosbag info <bag> (ROS 1) or ros2 bag info <bag> (ROS 2) to confirm the exact topic names, then update the config.

Running on a monocular camera

Set use_stereo:=false and max_cameras:=1 in the launch command. The EuRoC config uses both cameras by default; monocular mode works but produces less accurate depth estimates.

System runs slower than real-time

Reduce num_pts (e.g., to 100) and num_slam (e.g., to 25). You can also switch from the subscriber node to the serial node (serial.launch) which processes one frame at a time and is immune to scheduling delays.

Next steps

Try other EuRoC sequences (V1_02_medium, MH_01_easy, etc.) by downloading additional bags and passing the same config:=euroc_mav argument.
Explore other supported datasets (TUM-VI, UZH-FPV, KAIST Urban) in the Datasets guide.
Evaluate trajectory accuracy quantitatively using ov_eval — see the Evaluation overview.
Calibrate your own camera-IMU system following the Calibration guide.

Get Started

Guides

Theory

Evaluation

Development

Tutorial: Run OpenVINS on the EuRoC MAV Dataset

Configuration files

Launch files

Download the EuRoC bag

Launch OpenVINS

Visualize in RViz

Configuration overview

Tuning tips

Next steps

Build docs developers (and LLMs) love

Get Started

Guides

Theory

Evaluation

Development

Documentation Index

​Configuration files

​Launch files

​Download the EuRoC bag

​Launch OpenVINS

​Visualize in RViz

​Configuration overview

​Tuning tips

​Next steps

Build docs developers (and LLMs) love

Configuration files

Launch files

Download the EuRoC bag

Launch OpenVINS

Visualize in RViz

Configuration overview

Tuning tips

Next steps