Griffin: Aerial-Ground Cooperative 3D Perception Dataset with 200+ Scenes

Griffin is a pioneering publicly available dataset for aerial-ground cooperative 3D perception. It features over 200 dynamic scenes, totaling more than 30,000 frames and 270,000 images, built using CARLA-AirSim co-simulation. The dataset was created by author wjh-svm and last updated on Hugging Face in September 2025.

Use Cases

• Train 3D object detection models based on cooperative aerial-ground sensor data.
• Benchmark multi-agent tracking algorithms based on the dataset's dynamic scenes.
• Study occlusion effects in perception systems based on the instance-aware occlusion quantification.
• Develop sensor fusion methods based on data from variable UAV altitudes (20–60 meters).

Strengths

• Contains over 200 dynamic scenes, providing a substantial scenario library.
• Includes more than 30,000 frames and 270,000 images for model training.
• Features instance-aware occlusion quantification, a specific technical attribute.
• Uses CARLA-AirSim co-simulation, suggesting realistic sensor and environment modeling.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• Data may reflect simulation bias inherent to the CARLA-AirSim source.

Provenance

Source

Built using CARLA-AirSim co-simulation.

Collection Method

Simulation-generated.

Time Range

null

Freshness

Last updated 2025-09-04 01:31:51; freshness should be verified.

Geography

null