: A Unified Large-Scale Dataset for
Grounding Aerial Geometric 3D Vision

ECCV 2026

Xiaoya Cheng1,* Rouwan Wu1,* Xinyi Liu1,* Zeyu Cui1,* Yan Liu2,* Na Zhao3 Yu Liu1 Maojun Zhang1 Shen Yan1,†
1National University of Defense Technology, Changsha, China      2National Key Laboratory of Advanced Guidance and Control Technology
3Singapore University of Technology and Design, Singapore
*Equal contribution    †Corresponding author

arXiv

Evaluation Benchmarks

Engine

Code

A world-scale UAV simulation engine with pixel-perfect RGB-D geometry for aerial spatial intelligence.
Overview of AirZoo dataset and benchmarks
377 Regions
|
22 Countries
|
1.2M+ Frames
|
2,386 km Flight Distance

Abstract

Despite rapid progress in data-driven 3D vision, aerial geometric 3D vision remains limited by the scarcity of large-scale, high-fidelity training data. AirZoo bridges this gap with a unified dataset and benchmark for UAV-based sensing. It combines a scalable generation pipeline built on world-scale photogrammetric 3D meshes, comprehensive scene diversity across 377 regions in 22 countries, and rich geometric annotations including pixel-aligned metric depth, camera intrinsics, and precise 6-DoF geo-referenced poses. Through aerial image retrieval, cross-view matching, and multi-view 3D reconstruction, AirZoo acts as a pre-training engine that improves state-of-the-art models such as MegaLoc, RoMa, VGGT, and Depth Anything 3 on real-world aerial benchmarks.

Dataset Construction

AirZoo construction pipeline and properties

AirZoo uses Cesium for Unreal to stream Cesium 3D Tiles into UE5, then drives a custom AirSim-Cesium-Unreal simulator over global terrains. The pipeline logs continuous UAV video sequences rather than isolated screenshots, producing synchronized RGB images, dense metric depth, calibrated intrinsics, and Earth-fixed 6-DoF poses. Weather and illumination are systematically varied along the same trajectories, encouraging models to learn geometry that survives appearance changes.

Scale and Geometry

Comprehensive statistics of the AirZoo dataset

The dataset contains over 1.2 million high-resolution frames at 1600 × 1200 pixels, collected from nearly 2,386 km of simulated flight trajectories. Each region is rendered under multiple weather and time-of-day conditions, while the camera envelope spans 0-800 m altitude and 10°-90° pitch angles, covering oblique-to-nadir UAV viewpoints.

Geometric verification of AirZoo depth and poses

Bidirectional projection checks report a 0.066% median relative depth error, with P90 at 0.174% and P95 at 0.380%.

We further use AirZoo ground-truth camera poses and metric depth to build 3D Gaussian Splatting models. Four representative results are shown below. For each displayed trajectory, the rendering PSNR exceeds 30 dB.

Sydney Opera House
Eiffel Tower
Colosseum
Goddess of Victory

AirZoo-Real Benchmark

AirZoo also evaluates real-flight transfer through AirZoo-Real, a collected benchmark with RTK-aligned UAV imagery. The reconstruction split includes 9,430 images from four areas captured across 06:00-08:00, 12:00-14:00, and 18:00-20:00, with extra 22:00-24:00 low-light flights for two scenes. The benchmark resources are available through the Evaluation Benchmarks link above.

AirZoo-Real examples for reconstruction evaluation

Evaluation Tracks

1Aerial Image Retrieval

Given a UAV query, the retrieval task finds the most relevant geo-tagged satellite tile under large viewpoint, scale, illumination, and seasonal changes. AirZoo fine-tuning improves top-rank retrieval quality, especially on the harder AirZoo-Real split with arbitrary UAV viewpoints.

Qualitative cross-view geo-localization comparisons

2Cross-view Matching

Matching orthophoto references to oblique UAV imagery is central to UAV pose estimation. AirZoo exposes RoMa to difficult orthophoto-to-oblique pairs with controlled overlap, improving correspondence quality under altitude, viewpoint, and weather variation.

Qualitative cross-view matching results on AirZoo-Real

3Multi-view 3D Reconstruction

UAV sequences introduce wide baselines, strong oblique views, and lighting changes that ground-view training data rarely covers. Fine-tuning VGGT and DA3 on AirZoo consistently improves reconstruction performance across synthetic and real aerial evaluations.

Qualitative multi-view 3D reconstruction results

Citation

@article{cheng2026airzoo,
  author  = {Cheng, Xiaoya and Wu, Rouwan and Liu, Xinyi and Cui, Zeyu and Liu, Yan and Zhao, Na and Liu, Yu and Zhang, Maojun and Yan, Shen},
  title   = {AirZoo: A Unified Large-Scale Dataset for Grounding Aerial Geometric 3D Vision},
  journal = {arXiv preprint arXiv:2604.26567},
  year    = {2026},
  url     = {https://arxiv.org/abs/2604.26567}
}

Acknowledgements

AirZoo is built with Cesium for Unreal, Unreal Engine 5, and AirSim. Map data is obtained through Cesium. This website template is borrowed from longvolcap.