Back to feed
News Story
机器之心
1 sources

ECCV 2026 | AirZoo: Turning Global 3D Maps into Drone Training Grounds, a Unified Benchmark for Aerial Geometric 3D Vision

National University of Defense Technology and partners have released AirZoo, a large-scale unified dataset and benchmark for aerial geometric 3D vision. Covering 22 countries and 377 weather-condition trajectories, it includes over 1.2 million frames of pixel-level RGB-D data with 6-DoF geolocation annotations, supporting aerial image retrieval, cross-view matching, and multi-view 3D reconstruction. Experiments show that fine-tuning mainstream models on AirZoo significantly improves performance on aerial geometric 3D vision tasks. The work has been accepted by ECCV 2026.

SynthePulse Insight · AI deep reading

AirZoo: A Scalable Pretraining Engine for Drone Vision with Global 3D Maps

Version 1 · 1 source

A team from the National University of Defense Technology introduces AirZoo, a unified benchmark for aerial geometric 3D vision covering 22 countries and 1.2 million frames. Fine-tuning mainstream models with synthetic data yields significant improvements in retrieval, matching, and reconstruction tasks.

  • AirZoo synthetic data covers 22 countries, 377 weather trajectories, with 1.2 million frames of pixel-level RGB-D data and 6-DoF pose annotations.
  • Fine-tuning MegaLoc with AirZoo improves R@1 from 6.46% to 18.66% on the real-world benchmark AirZoo-Real.
  • Fine-tuning RoMa reduces median cross-view matching error to 3.03 meters, with acc@5m reaching 79.78%.
  • Fine-tuning Depth Anything 3 boosts F1 from 54.10 to 86.09 on AirZoo-Test.
  • The AirZoo-Real benchmark covers areas like Changsha and provides RTK-accurate ground truth.
  • The research has been accepted by ECCV 2026.
Open section navigationThe Challenge of Aerial Geometric 3D Vision and AirZoo's Response

The Challenge of Aerial Geometric 3D Vision and AirZoo's Response

Existing visual foundation models perform well on ground-level scenes, but drone aerial imagery faces drastic changes in viewpoint, scale, and environment, leading to poor model transfer. Current UAV datasets either lack geometric supervision or cover only a few handcrafted scenes, making them insufficient for sequence-level geometric learning.

AirZoo's core idea is to integrate global-scale real geographic scenes, controllable drone motion, and pixel-level geometric supervision into an automated pipeline. Based on AirSim, Cesium, and Unreal Engine, it generates flight trajectories from WGS84 waypoints, synchronously rendering RGB, metric depth, and 6-DoF poses, outputting continuous 30fps trajectory-level data.

The data covers 22 countries across six continents, with 95 base sequences and 377 weather trajectories, totaling approximately 2,400 km of flight, altitudes from 0 to 800 meters, pitch angles from 10° to 90°, simulating sunny, cloudy, rainy, foggy, snowy, and day/night conditions.

Geometric Validation and Sim-to-Real Transfer

AirZoo provides pinhole camera intrinsics and 6-DoF poses in both WGS84 and ECEF coordinate systems for each frame. Bidirectional reprojection checks between adjacent frames show median relative depth error below 0.1%. The team replicates real DJI video viewpoints in AirZoo, and the synthetic views are highly comparable to real aerial footage in layout, scale, and texture.

The team releases the AirZoo-Real benchmark, covering urban and rural scenes in Changsha and other areas, with different day/night periods and RTK-accurate ground truth, to evaluate algorithm performance in real-world conditions.

Significant Improvements Across Three Tasks

In aerial image retrieval, fine-tuning MegaLoc with AirZoo improves R@1 from 6.46% to 18.66% and R@5 from 22.16% to 50.53% on AirZoo-Real. Qualitative results show that the fine-tuned model retrieves correct regions more stably under complex viewpoints and lighting.

In cross-view matching, fine-tuning RoMa reduces median error to 3.03 meters, with acc@5m reaching 79.78%. Matching points are more evenly distributed on static structures like roads and buildings, with significantly fewer false correspondences.

In multi-view 3D reconstruction, fine-tuning Depth Anything 3 improves F1 from 54.10 to 86.09 on AirZoo-Test; fine-tuning VGGT improves F1 from 43.38 to 52.61 on UrbanScene3D. Reconstructions are more complete with fewer artifacts.

Summary and Outlook

AirZoo integrates globally scalable simulation, pixel-level geometric ground truth, and unified evaluation across three tasks into a platform for drone spatial intelligence. Experiments show that large-scale, diverse synthetic data with precise geometric supervision effectively transfers visual foundation models to real drone tasks.

The team plans to further expand data scale and real benchmark coverage, promoting the deployment of aerial geometric 3D vision in smart cities, emergency mapping, and unmanned systems.

Credibility boundary

This article is primarily based on a report by Jiqizhixin on the work of the National University of Defense Technology team. The paper has been accepted by ECCV 2026, and the data is publicly available. All experimental results are from the paper and have not been independently verified.

Insight takeaway

AirZoo provides a scalable pretraining solution for aerial geometric 3D vision using global-scale synthetic data. Fine-tuning mainstream models yields significant improvements in retrieval, matching, and reconstruction tasks, demonstrating the potential of synthetic data for transfer to real-world scenarios.

Sources for this version

  1. ECCV 2026｜把全球3D地图变成无人机训练场，国防科大AirZoo构建航拍几何3D视觉统一基准机器之心

Primary report

机器之心

Primary source