Skip to content

Dataset

The IILABS 3D dataset was collected in the Industry and Innovation Laboratory. The dataset is composed by calibration and benchmark sequences. Each sequence contains data from multiple sensors, complemented by high-precision ground truth obtained via a Motion Capture (MoCap) system.

Dataset Overview

Key Features

  • Multiple 3D LiDAR Sensors: Data from four distinct 3D LiDAR sensors with different characteristics
  • Additional Sensors: 2D LiDAR, IMU, and wheel odometry data
  • Diverse Sequences: Both calibration sequences and challenging benchmark trajectories
  • High-Precision Ground Truth: Sub-millimeter accuracy from a 24-camera OptiTrack system

Sensors

The dataset includes data from the following sensors:

3D LiDARs

  • Livox Mid-360: Solid-state LiDAR with an non-repetitive scanning pattern
  • Ouster OS1-64 RevC: 64-channel mechanical spinning LiDAR (45° uniform vertical FoV)
  • RoboSense RS-HELIOS-5515: 32-channel mechanical spinning LiDAR (70° non-uniform vertical FoV)
  • Velodyne VLP-16: 16-channel mechanical spinning LiDAR (30° uniform vertical FoV)

Additional Sensors

  • Hokuyo UST-10LX: 2D LiDAR
  • Xsens MTi-630 AHRS: Inertial Measurement Unit (IMU)
  • Faulhaber 2342 wheel encoders: 64:1 gear ratio, 12 Counts Per Revolution (CPR)

Sensor Details

For detailed specifications of each sensor, please refer to the Sensors section.

Data Formats

The dataset provides three types of data:

  1. Sensor Data: Provided in ROS 1 bag files, with each file representing a sequence during which the robot followed a defined trajectory.

  2. Ground Truth Data: Available as text files in TUM format (8 columns: timestamp in seconds, three columns for position coordinates, and four columns for orientation in quaternion format).

  3. Calibration Files: Supplied in YAML format, containing the calibration parameters for each sensor.

Convert dataset to ROS 2

Even though the dataset is provided in ROS 1 bag files, we provide a way to convert to ROS 2 bag formats in our IILABS 3D toolkit.

ROS Topics and Explanation of Each Message

Topic Name
 Message Type Description Frequency
/eve/lidar3d sensor_msgs/PointCloud2 raw pointcloud data (Velodyne VLP-16 / RoboSense RS-Helios-5515 / Livox Mid-360) 10Hz
/eve/ouster/points sensor_msgs/PointCloud2 raw pointcloud data (Ouster OS1-64) 10Hz
/eve/imu/data sensor_msgs/Imu raw IMU data (Xsens MTi-630) 400Hz
/eve/ouster/imu sensor_msgs/Imu raw IMU data (Ouster OS1-64) 100Hz
/eve/livox/imu sensor_msgs/Imu raw IMU data (Livox Mid-360) 200Hz
/eve/scan sensor_msgs/LaserScan raw laser scan data (Hokuyo UST-10LX) 40Hz
/eve/odom nav_msgs/Odometry wheel odometry processed data 100Hz
/eve/motors_enc sdpo_drivers_interfaces/MotEncArrayROS1 raw motor encoders data 100Hz
/eve/motors_ref sdpo_drivers_interfaces/MotRefArrayROS1 motor reference speeds -
/tf tf2_msgs/TFMessage reference frame transformations -
/tf_static tf2_msgs/TFMessage static reference frame transformations -

Custom ROS Messages

The messages for the motor encoders data (sdpo_drivers_interfaces/MotEncArrayROS1) and motor reference speeds (sdpo_drivers_interfaces/MotRefArrayROS1) are custom ROS messages. These message definitions are included in this repository as a Git submodule via the 5dpo_drivers_interfaces package, which supports both ROS 1 and ROS 2. If you are cloning this repository, make sure to also initialize and update the submodules: 

git submodule update --init --recursive
Moreover, when converting the dataset ROS bags to ROS 2 using the IILABS 3D Toolkit, the message type of these topics is automatically adjusted for the ROS 2 equivalent.

Data Collection Method

Sensor data was captured using the Robot Operating System (ROS) framework's rosbag record tool on a LattePanda 3 Delta embedded computer. Post-processing involved timestamp correction for the Xsens MTi-630 AHRS IMU via custom Python scripts.

Ground-truth data was captured using an OptiTrack MoCap system featuring 24 high-resolution PrimeX 22 cameras. These cameras were connected via Ethernet to a primary Windows computer running the Motive software, which processed the camera data. This Windows computer was then connected via Ethernet to a secondary Ubuntu machine running the NatNet 4 ROS driver. The bag files were processed using the EVO open-source Python library to convert the data into TUM format and adjust the initial position offsets for accurate SLAM odometry benchmarking.

Temporal synchronization between the robot platform and the ground-truth system was achieved using the Network Time Protocol (NTP).

Accessing the Dataset

The IILABS 3D dataset is available at the INESC TEC research data repository. The dataset is approximately 350 GB in size.

Using the IILABS 3D Toolkit

The easiest way to access and work with the dataset is through the IILABS 3D Toolkit, a Python package that provides utilities for working with the dataset.

# Install the toolkit
pip install iilabs3d-toolkit

# List available sequences
iilabs3d list-sequences

# List available sensors
iilabs3d list-sensors

# Download a specific sequence for a specific sensor
iilabs3d download ~/slam_data loop livox_mid_360

Downloading All Data

To download all benchmark sequences for all sensors: 

iilabs3d download ~/slam_data bench all

Additional Information

For more information on using the toolkit, see the IILABS 3D Toolkit GitHub page.