Griffin: Aerial-Ground Cooperative Detection and Tracking Dataset and Benchmark

Griffin is the pioneering publicly available dataset for aerial-ground cooperative 3D perception. Built using CARLA-AirSim co-simulation, it features over 200 dynamic scenes—totaling more than 30,000 frames and 270,000 images. With instance-aware occlusion quantification, variable UAV altitudes (20–60 meters), and realistic drone dynamics under diverse conditions, Griffin sets a new standard for cooperative perception research.

---

News

• 🚀 (2025/3) Released source code and pre-trained models for the AGILE framework—the first Aerial-Ground Instance-LEvel intermediate fusion framework.
• 📦 (2025/3) Griffin V1.0 dataset is now available at Baidu Netdisk.

---

Getting Started

• Installation
• Dataset Preparation
• Training and Evaluation
• Visualization

---

Main Results

Detection and tracking on Griffin-25m under various latency conditions

Fusion Stage	Method	Latency (ms)	AP ↑	ATE ↓	AOE ↓	AMOTA ↑	AMOTP ↓	MT ↑	ML ↓	IDS ↓	Comm. Cost ↓
No Fusion	-	0	0.366	0.398	0.492	0.363	1.295	16	49	5	0
Early Fusion	Concat	0	0.626	0.385	0.795	0.653	0.869	34	16	38	~3×10⁸
		200	0.510	0.502	0.776	0.561	1.044	25	21	34
		400	0.416	0.522	0.714	0.418	1.189	19	32	26
Instance Fusion	AGILE	0	0.412	0.484	0.708	0.435	1.130	21	35	23	~6×10⁵
		200	0.396	0.484	0.705	0.410	1.165	16	36	20
		400	0.373	0.505	0.709	0.366	1.212	15	40	12
Late Fusion	Hungarian	0	0.377	0.357	0.511	0.369	1.078	14	55	16	~1×10⁴
		200	0.363	0.362	0.498	0.357	1.060	15	55	11
		400	0.339	0.350	0.458	0.310	1.129	15	58	11

Performance Across UAV Altitudes

Dataset	Fusion Stage	AP	AMOTA
Griffin-25m	No Fusion	0.366	0.363
	Early Fusion	0.626 (+71.0%)	0.653 (+79.9%)
	Instance Fusion	0.412 (+12.6%)	0.435 (+19.8%)
	Late Fusion	0.377 (+3.0%)	0.369 (+1.7%)
Griffin-40m	No Fusion	0.351	0.371
	Early Fusion	0.495 (+41.0%)	0.543 (+46.4%)
	Instance Fusion	0.359 (+2.3%)	0.397 (+7.0%)
	Late Fusion	0.354 (+0.9%)	0.387 (+4.3%)
Griffin-Random	No Fusion	0.465	0.491
	Early Fusion	0.580 (+24.7%)	0.646 (+31.6%)
	Instance Fusion	0.408 (-12.3%)	0.426 (-13.2%)
	Late Fusion	0.375 (-19.4%)	0.394 (-19.8%)

Detailed results with all metrics are lised in docs/detailed_results.csv.

---

Citation

If you find Griffin useful, please consider giving a ⭐ and citing our work (details coming soon): TBD

---

Acknowledgement

We're grateful to these amazing open-source projects:

• mmdet3d: Core 3D detection framework.
• UniV2X: Cooperative perception inspiration (by our team).
• BEVFormer: 3D Object Detection Baseline.
• AB3DMOT: 3D Multi-Object Tracking Baseline.