Turning Global 3D Maps into a UAV Training Ground: AirZoo’s Unified 3D Vision Benchmark

The SAW Lab at the National University of Defense Technology released AirZoo, a unified large‑scale dataset and benchmark for aerial geometric 3D vision. The team built an automated pipeline based on AirSim, Cesium for Unreal, and Unreal Engine that generates continuous 30 fps RGB‑D streams, pixel‑level depth, camera intrinsics, and 6‑DoF poses in both WGS84 and ECEF coordinates from globally distributed WGS84 waypoints.

AirZoo covers 22 countries across six continents, with 95 base flight sequences and 377 weather‑condition trajectories, totaling about 2 400 km of flight and more than 1.2 million frames at 1600×1200 resolution. The data simulate sunny, cloudy, rainy, foggy, snowy, daytime, dusk, and night conditions, and camera tilt angles sweep from 10° to 90° while altitude ranges from 0 to 800 m.

Each frame includes precise geometric supervision: camera intrinsics, dual‑coordinate 6‑DoF poses, and a bidirectional reprojection check that yields a median relative depth error below 0.1 %. Synthetic views were compared with real DJI footage captured at identical viewpoints, showing high fidelity in scene layout, scale, and texture.

The dataset unifies three core aerial 3D‑vision tasks—image retrieval, cross‑view matching, and multi‑view 3D reconstruction—using the same training and test splits (AirZoo‑Real provides real‑flight evaluation with RTK‑level positioning). Experiments kept model architectures unchanged and fine‑tuned them on AirZoo.

For image retrieval, MegaLoc fine‑tuned on AirZoo improved R@1 on AirZoo‑Real from 6.46 % to 18.66 % and R@5 from 22.16 % to 50.53 %, demonstrating more stable geographic localization under extreme view and illumination changes.

In cross‑view matching, RoMa fine‑tuned on AirZoo reduced median error on AirZoo‑Real to 3.03 m and raised acc@5 m to 79.78 %, with matches distributed across static structures rather than erroneous texture‑based correspondences.

For multi‑view 3D reconstruction, Depth Anything 3’s F1 score on AirZoo‑Test rose from 54.10 to 86.09, and VGGT’s F1 on UrbanScene3D increased from 43.38 to 52.61. Qualitative results show fewer missing structures, reduced geometric breaks, and fewer artifacts after AirZoo fine‑tuning.

Overall, AirZoo provides a scalable, controllable, and geometrically verified data engine that bridges the gap between synthetic simulation and real UAV deployments, enabling vision foundation models to acquire aerial geometric priors and improve transfer to real‑world tasks such as smart‑city mapping, emergency surveying, and autonomous UAV operations.