AirZoo releases a global aerial 3D vision dataset with 1.2 million RGB-D frames
机器之心 · wechat · 2026-07-23
AirZoo unifies aerial geometric 3D vision with a global-scale dataset and benchmark
The NUDT SAWLab team released AirZoo, a large-scale dataset and benchmark for aerial geometric 3D vision, along with an automated AirSim-Cesium-Unreal simulation pipeline.
- Scale: 22 countries across 6 continents, 377 weather-conditioned trajectories, over 1.2 million RGB-D frames, and about 2,400 km of simulated flight.
- Coverage: cities, rural areas, and natural landscapes, with conditions spanning sun, overcast, rain, fog, snow, day, dusk, and night.
- Tasks: aerial image retrieval, cross-view matching, and multi-view 3D reconstruction.
- Validation: the team also released AirZoo-Real, a real-flight evaluation set with RTK positioning to test sim-to-real transfer.
- Results: fine-tuning existing models on AirZoo significantly improved performance on both public benchmarks and AirZoo-Real, suggesting the dataset fills an important domain gap for UAV spatial intelligence.
More from Research
- Sierra’s internal agent hit the same bottleneck: getting the right context — blaizedsouza · 2026-07-23
- Paper argues algorithmic neutrality is impossible because relevance itself is value-laden — CurieuxExplorer · 2026-07-23
- Anchor-Align lifts xArm7 real-robot success from 28% to 54% — Dwip Dalal · 2026-07-23
- SetwiseEvalKit scores document sets, not just single search results — Kailin Jiang · 2026-07-23
- DeepSeek-V4 post-training on Ascend SuperPOD reaches 34.22% MFU — Dongfang Li · 2026-07-23
- Essay says “algorithmic neutrality” is a myth and warns philosophy hires can become corporate ethics theater — CurieuxExplorer · 2026-07-23