---

🎯 What is Griffin?

Griffin is a pioneering publicly large-scale dataset specifically designed for aerial-ground cooperative 3D perception. Our dataset pushes the boundaries of multi-agent perception by combining aerial and ground-based viewpoints for enhanced 3D object detection and tracking.

✨ Key Features

• 🎬 250+ Dynamic Scenes with realistic traffic patterns
• 📸 37K Frames and 340K High-quality Images
• 🎯 Instance-aware Occlusion Analysis for precise labels
• ✈️ Multi-altitude Drone Simulation (20m-60m)
• 🌍 CARLA-AirSim Co-simulation for photorealistic environments
• 🎪 Comprehensive Benchmarks for detection and tracking
• 📡 Robustness Evaluation under communication interference and localization errors

📋 Table of Contents

• 🦅 Griffin
  • 🎯 What is Griffin?
    • ✨ Key Features
  • 📋 Table of Contents
  • 🔥 Latest News
  • 📚 Documentation
  • 📈 Main Results
    • 🎯 Baseline Performance
    • 🌐 Communication Robustness
    • 📍 Localization Robustness
    • 🏆 Key Insights
  • 📝 Citation
  • 🙏 Acknowledgements

---

🔥 Latest News

🚨 Stay updated with the latest developments in Griffin!

Date	Update	Description
2025/8	🔧 Robustness Evaluation	Testing configurations for localization errors, communication latency, and packet loss are now available
2025/7	📊 Griffin-55m Subset	New subset Griffin-55m and corresponding model checkpoints are released
2025/3	🤖 UniV2X Models	Released reimplementation code and pre-trained models for UniV2X
2025/3	💾 Dataset V1.0	Griffin V1.0 dataset is available on Baidu Netdisk and 🤗 Hugging Face
2025/3	📄 Paper Published	Our paper is now available on ArXiv

---

📚 Documentation

Comprehensive guides to help you get the most out of Griffin:

Guide	Description	Link
🛠️ Installation	Step-by-step setup instructions	docs/Installation.md
📊 Dataset Preparation	How to download and organize the data	docs/Dataset_Preparation.md
🏃‍♂️ Training & Evaluation	Run experiments and evaluate models	docs/Training_and_Evaluation.md
🎨 Visualization	Visualize results and debug your models	docs/Visualization.md

---

📈 Main Results

Griffin provides comprehensive benchmarks across multiple models and challenging scenarios. Our evaluation covers detection and multi-object tracking metrics under various conditions.

🎯 Baseline Performance

The AP and AMOTA metrics of every baseline among different subsets are shown below. For detailed results with all metrics, see 📊 detailed_results.csv.

🌐 Communication Robustness

📍 Localization Robustness

🏆 Key Insights

• 🤝 Cooperative Potential: In favorable conditions, cooperative methods achieve substantial performance gains over single-agent baselines by resolving occlusions and expanding the effective field-of-view
• ✈️ Altitude Sensitivity: Strong sensitivity to drone flight altitude affects performance, with instance-level fusion strategies proving more resilient to perspective shifts than dense BEV-level approaches
• 🎯 Adaptive Filtering: Resilience to localization errors is directly linked to adaptive data filtering—methods with selective fusion (instance-level filtering or spatial confidence maps) demonstrate superior robustness
• 🔮 Future Directions: Research should focus on altitude-adaptive fusion mechanisms, sparse communication-efficient methods, and dynamic trust mechanisms for reliable real-world deployment

---

📝 Citation

If you find Griffin useful for your research, please consider giving us a ⭐ and citing our work:

@misc{wang2025griffinaerialgroundcooperativedetection,
      title={Griffin: Aerial-Ground Cooperative Detection and Tracking Dataset and Benchmark},
      author={Jiahao Wang and Xiangyu Cao and Jiaru Zhong and Yuner Zhang and Haibao Yu and Lei He and Shaobing Xu},
      year={2025},
      eprint={2503.06983},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2503.06983},
}

---

🙏 Acknowledgements

We extend our heartfelt gratitude to the amazing open-source community and these outstanding projects that made Griffin possible:

---