Reproduce “AirZoo: A Unified Large-Scale Dataset for Grounding Aerial Geometric 3D Vision” on OpenResearch

About this paper

Xiaoya Cheng, Rouwan Wu, Xinyi Liu, Zeyu Cui, Yan Liu, Na Zhao, Yu Liu, Maojun Zhang and 1 more

AirZoo is a new large-scale synthetic dataset containing over 1.2 million high-resolution aerial frames with precise geometric annotations, designed to train 3D vision models for unmanned aerial vehicles (UAVs). Generated using a custom AirSim-Cesium-Unreal simulator across 22 countries, the dataset improves state-of-the-art models' performance on aerial image retrieval, cross-view matching, and multi-view 3D reconstruction tasks.

Key insights

• Large-scale synthetic data, when designed to match aerial conditions and provide rich geometric supervision, can effectively bridge the domain gap between terrestrial-trained models and UAV applications.
• The inclusion of pixel-level metric depth maps and precise 6-DoF poses is crucial for enabling models to learn robust, geometry-aware features for aerial tasks.
• While AirZoo training consistently improves performance, a domain gap persists between synthetic and real-world results, indicating ongoing challenges for real-flight trajectories.
• Specific flight patterns, if not represented in the training data (e.g., circular trajectories around buildings), can lead to model failures, highlighting the importance of trajectory diversity.

Summary by alphaXiv.