AirZoo-Real

AirZoo-Real is a real-world UAV benchmark introduced in our ECCV 2026 paper, AirZoo: A Unified Large-Scale Dataset for Grounding Aerial Geometric 3D Vision.

This repository contains two real survey scenes and an additional reconstruction-oriented release. The data supports aerial image retrieval, cross-view matching, visual localization, 3D reconstruction, and geometry-aware evaluation. It includes UAV imagery, positioning records, digital orthophoto maps (DOM), digital surface/elevation models, rendered reconstruction images, camera intrinsics, camera poses, and timestamps.

Paper

AirZoo: A Unified Large-Scale Dataset for Grounding Aerial Geometric 3D Vision
Xiaoya Cheng, Rouwan Wu, Xinyi Liu, Zeyu Cui, Yan Liu, Na Zhao, Yu Liu, Maojun Zhang, and Shen Yan
Accepted at the European Conference on Computer Vision (ECCV), 2026.

Dataset summary

Directory	Content	Images	Files	Size	Notes
Scene1	Urban, building-dense UAV survey scene	161 UAV query images	173	3.49 GiB	CRS: EPSG:4547
Scene2	Rural, farmland-dominated UAV survey scene	190 UAV query images	201	2.82 GiB	CRS: EPSG:4547
airzoo_real_recon	Multi-site reconstruction data with rendered images, poses, intrinsics, timestamps, and DOM/DEM rasters	28,292 PNG images	28,377	29.27 GiB	Four sites plus shared DOM/DEM data
Total	Full local release	28,643 images	28,751	35.58 GiB	Scene data plus reconstruction data

Directory structure

AirZoo-Real/
|-- README.md
|-- Scene1/
|   |-- query/
|   |   |-- *.JPG
|   |   |-- *.MRK
|   |   |-- *.NAV
|   |   |-- *.OBS
|   |   |-- *.RTK
|   |   `-- ppk_file_name.pbk
|   |-- dom_dsm/
|   |   |-- DOM.tif
|   |   |-- DOM.tfw
|   |   |-- DSM.tif
|   |   |-- DSM.tfw
|   |   `-- metadata.xml
|   `-- satemap/
|       `-- dom.tif
|-- Scene2/
|   `-- ...
`-- airzoo_real_recon/
    |-- biandian/
    |   |-- 06-08/
    |   |   |-- images/
    |   |   `-- poses/
    |   |       |-- intrinsics/t_intrinsic.txt
    |   |       |-- poses/t_pose.txt
    |   |       `-- times/
    |   |-- 12-14/
    |   `-- 18-20/
    |-- guangchang/
    |   |-- 06-08/
    |   |-- 12-14/
    |   `-- 18-20/
    |-- jiaxiao/
    |   |-- 06-08/
    |   |-- 12-14/
    |   |-- 18-20/
    |   `-- 22-24/
    |-- xuexiao/
    |   |-- 06-08/
    |   |-- 12-14/
    |   |-- 18-20/
    |   `-- 22-24/
    `-- dom/
        |-- fcw_hangtian_DOM.tif
        |-- fcw_hangtian_DSM.tif
        |-- fcw_hangtian_DSM_cached.npy
        `-- allfeicuiwan/
            |-- DOM.tif
            |-- DOM_wgs84.tif
            `-- DEM/

Data fields

For Scene1 and Scene2:

• query/: oblique UAV images and associated DJI/PPK positioning records.
• dom_dsm/DOM.tif: high-resolution digital orthophoto map.
• dom_dsm/DSM.tif: digital surface model aligned with the DOM.
• dom_dsm/*.tfw: world files for the corresponding GeoTIFF rasters.
• dom_dsm/metadata.xml: projected coordinate reference metadata.
• satemap/dom.tif: satellite reference map used for cross-view evaluation.

For airzoo_real_recon:

• images/: rendered PNG images for each site and split.
• poses/intrinsics/t_intrinsic.txt: camera intrinsic parameters.
• poses/poses/t_pose.txt: camera pose records.
• poses/times/: timestamp arrays and text exports.
• dom/: shared DOM/DEM rasters and cached elevation data.

Download

Download the complete repository with Hugging Face tooling:

pip install -U huggingface_hub
hf download RingoWRW97/AirZoo-Real \
  --repo-type dataset \
  --local-dir AirZoo-Real

Or download it from Python:

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="RingoWRW97/AirZoo-Real",
    repo_type="dataset",
    local_dir="AirZoo-Real",
)

To download only part of the repository:

snapshot_download(
    repo_id="RingoWRW97/AirZoo-Real",
    repo_type="dataset",
    local_dir="AirZoo-Real",
    allow_patterns=["airzoo_real_recon/**"],
)

Replace airzoo_real_recon/** with Scene1/** or Scene2/** to download only one survey scene.

Notes

• The repository contains large GeoTIFF files; use geospatial software that can read tiled rasters without loading an entire raster into memory.
• Raw flight, pose, timestamp, RTK, and geospatial metadata may contain precise location information.
• The repository stores source files directly and does not define a Hugging Face datasets loading script.

Citation

If you use AirZoo-Real in your research, please cite our ECCV 2026 paper:

@inproceedings{cheng2026airzoo,
  title     = {{AirZoo}: A Unified Large-Scale Dataset for Grounding Aerial Geometric {3D} Vision},
  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},
  booktitle = {European Conference on Computer Vision (ECCV)},
  year      = {2026}
}

License

Dataset licensing terms will be added by the authors before public release. Until then, do not assume permission to redistribute or use the data beyond the terms communicated by the dataset owners.