license: cc-by-4.0
size_categories:

• 1K<n<10K

UE4-Stereo

Dataset Description

UE4-Stereo is a synthetic stereo vision dataset generated in Unreal Engine 4.27 using the Microsoft AirSim simulation framework.
The dataset provides synchronized stereo RGB images, dense ground-truth depth maps, and full 6-DoF camera poses. It is designed
to support research and benchmarking in stereo matching, visual odometry, and visual SLAM, particularly in controlled indoor environments.

The dataset focuses on realistic indoor scenes with structural elements such as walls, pipes, pillars, and furniture,
while allowing precise control over camera configuration and motion.

Dataset Overview

• Number of scenes: 2 indoor scenes
• Number of sequences: 16 acquisition sequences (8 per scene)
• Total stereo image pairs: 9,810
• Trajectory type: Closed-loop trajectories
• Sampling frequency: 3 Hz
• Drone velocity: 1.5 m/s

Each sequence corresponds to a complete loop, where the drone takes off, follows a predefined trajectory, and lands at the same initial position.

Acquisition Setup

• Simulation engine: Unreal Engine 4.27
• Simulation framework: Microsoft AirSim
• Platform: Virtual drone
• Cameras: Stereo RGB cameras
• Image resolution: 640 × 480 pixels
• Camera intrinsics: Fixed across all sequences
• Stereo baseline: 24 cm and 32 cm (depending on the sequence)
• Camera configuration: Parallel and convergent (±5°) setups

Drone motion and data acquisition were controlled via Python scripts using the AirSim API. All data streams are temporally synchronized and referenced to the left camera.

Scenes and Sequences

Scene 1

Scene 1 represents an indoor industrial-like environment. It contains eight sequences with controlled variations in:

• Loop direction (clockwise and counter-clockwise)
• Lateral drone displacement (~1 m)
• Stereo baseline (24 cm and 32 cm)
• Camera convergence (parallel and ±5° convergent)

The camera orientation remains fixed across all sequences in this scene.

Scene 2

Scene 2 represents a second indoor environment with similar structural complexity. It also contains eight sequences with variations in:

• Loop direction (clockwise and counter-clockwise)
• Lateral drone displacement (~1 m)
• Stereo baseline (24 cm and 32 cm)
• Camera convergence (parallel and ±5° convergent)

Data Organization

The dataset is organized as follows:

scene_01/
readme_scene1.txt
sequence01/
left_rgb/
right_rgb/
depth_npy/
depth_visual/
associate_rgb.txt
associate_rgbd.txt
groundtruth.txt
sequence02/
...
...
sequence08/

scene_02/
readme_scene2.txt
sequence01/
left_rgb/
right_rgb/
depth_npy/
depth_visual/
associate_rgb.txt
associate_rgbd.txt
groundtruth.txt
...
sequence08/
README.txt

Data Formats

RGB Images

• Format: PNG
• Resolution: 640 × 480
• Channels: RGB (8-bit)
• Stored separately for left and right cameras

Depth Maps

• Metric depth: NumPy arrays (.npy)

  • Data type: float32
  • Units: meters
  • Resolution: 640 × 480
  • Pixel-wise aligned with the left RGB image

• Depth visualization: PNG images

  • 8-bit grayscale
  • Normalized for visualization only
  • Not suitable for quantitative evaluation

Camera Poses

• Format: Text (groundtruth.txt)
• Pose representation:

timestamp tx ty tz qx qy qz qw

• Position: (tx, ty, tz) in meters
• Orientation: quaternion (qx, qy, qz, qw)
• Coordinate system: Unreal Engine world frame

The pose format follows the standard convention used by the TUM RGB-D dataset.

Association Files

Each sequence provides:

• associate_rgb.txt: Synchronization between left and right RGB images
• associate_rgbd.txt: Synchronization between left RGB images and depth maps

These files guarantee one-to-one correspondence between all data modalities.

Intended Use

This dataset is intended for:

• Stereo disparity estimation
• Visual odometry
• Visual SLAM
• Domain adaptation and synthetic-to-real transfer
• Evaluation of robustness to camera geometry variations

Note: This is not a classification dataset. Folder names and file organization should not be interpreted as class labels.

License

This dataset is released under the Creative Commons Attribution 4.0 (CC BY 4.0) license.

Citation

If you use this dataset, please cite the corresponding Data in Brief article (under review).
The citation will be updated upon acceptance.