The Dataset Viewer has been disabled on this dataset.

Songdo Vision: Vehicle Annotations from High-Altitude BEV Drone Imagery in a Smart City

Zenodo GitHub License Paper arXiv Geo-trax Model Demo Space Songdo Traffic Songdo Vision OBB Stabilo Benchmark Collection Website

Songdo Vision is a high-resolution aerial vehicle-detection dataset of 4K bird's-eye view (BEV) RGB images with axis-aligned bounding-box annotations. It comprises 5,419 annotated drone frames containing ~274k vehicle instances across four classes (car, bus, truck, motorcycle), captured during a large-scale multi-drone urban traffic monitoring experiment over the Songdo International Business District, South Korea. It is the detection dataset behind the Geo-trax pipeline and the associated publication.

Highlights

  • πŸ–ΌοΈ Native 4K, true top-down frames from 140–150 m: small, densely packed vehicles that ground-level and oblique datasets do not cover.
  • βœ… 274,190 human-verified boxes; in v2 every removal and addition was confirmed by a human, and all three formats (COCO, YOLO, Pascal VOC) are generated from one source and agree exactly.
  • πŸ‹οΈ Proven in training: the Geo-trax detector trained with it reaches 0.951 mAP@50; try it in the live demo.
  • πŸ”„ Also available with oriented boxes, and paired with ~700k trajectories from the same experiment.

πŸ“¦ The data is hosted on Zenodo, not on Hugging Face. This page is a documentation mirror and index card. The full dataset (images + annotations, ~16.1 GB) lives immutably on Zenodo under DOI 10.5281/zenodo.13828407. See Access the dataset below for how to download it.

πŸ”„ Need oriented boxes? The same frames are also available with rotated (oriented) bounding boxes, converted from the v2 annotations, as Songdo Vision OBB (10.5281/zenodo.15050578).

Dataset at a glance

Property Value
Modality 4K RGB images, bird's-eye view (BEV) β€” 3840 Γ— 2160, except 8 frames recorded portrait (2160 Γ— 3840)
Frames 5,419 (4,335 train / 1,084 test; 80/20 split)
Instances 274,190 (218,662 train / 55,528 test)
Classes 4: 0 car (incl. vans), 1 bus, 2 truck, 3 motorcycle
Annotation type Axis-aligned (horizontal) bounding boxes
Formats COCO JSON, YOLO TXT (normalized), Pascal VOC XML
Acquisition Fleet of 10 drones over 20 intersections (primarily) and some roads in between, 29.97 FPS, 140–150 m altitude, DJI Mavic 3
Collection dates October 4–7, 2022
Location Songdo International Business District, South Korea
Total size ~16.1 GB (train.zip + test.zip)
License CC BY 4.0

Class distribution

ID Class Train Test Total
0 Car (incl. vans) 196,719 49,862 246,581
1 Bus 7,052 1,766 8,818
2 Truck 11,861 3,080 14,941
3 Motorcycle 3,030 820 3,850
Total 218,662 55,528 274,190

What changed in v2

v2 is an annotation-quality release; the images are byte-identical to v1. 196 invalid boxes were removed (road markings, rooftop plant, construction machinery, duplicates, degenerate boxes) and 1,951 vehicles missed by the original annotation pass were added. All three annotation formats are now generated from a single source and agree exactly, where COCO and Pascal VOC previously disagreed by up to 1 px on ~75% of boxes.

Class v1 Removed Added v2
Car 245,047 140 1,674 246,581
Bus 8,789 10 39 8,818
Truck 14,831 34 144 14,941
Motorcycle 3,768 12 94 3,850
Total 272,435 196 1,951 274,190

Candidates for both removals and additions were ranked with a detector and then confirmed by a human reviewer; nothing was changed automatically. Annotations carried over from v1 keep their original COCO ids (≀ 272435); annotations added in v2 are numbered from 1000001, so the two sets can be told apart β€” filtering id <= 272435 reproduces the v1-provenance annotation set. See the dataset README on Zenodo for the full changelog and known limitations.

Access the dataset

The dataset is published and versioned on Zenodo. Please download it from the canonical source; it is not redistributed on Hugging Face:

➑️ DOI 10.5281/zenodo.13828407 (concept DOI β€” always resolves to the latest version, currently v2, version DOI 10.5281/zenodo.21728430; v1 is 10.5281/zenodo.13828408)

Files on Zenodo:

File Size Contents
train.zip β‰ˆ12.9 GB Training images + annotations (COCO / YOLO / Pascal VOC)
test.zip β‰ˆ3.2 GB Test images + annotations (COCO / YOLO / Pascal VOC)
names.txt small Class names
data.yaml small Example YOLO dataset configuration
README.md small Dataset documentation
LICENSE.txt small CC BY 4.0 terms

Download with zenodo_get:

pip install zenodo_get
zenodo_get 10.5281/zenodo.13828407      # fetches all files for the latest version

Or directly with requests:

import requests

# Follow the concept DOI to whichever record is currently the latest version,
# then download that record's files.
record_id = requests.get(
    "https://doi.org/10.5281/zenodo.13828407", allow_redirects=True
).url.rstrip("/").split("/")[-1]
record = requests.get(f"https://zenodo.org/api/records/{record_id}").json()
for f in record["files"]:
    url = f["links"]["self"]
    print("Downloading", f["key"], f"({f['size'] / 1e9:.2f} GB)")
    with requests.get(url, stream=True) as r, open(f["key"], "wb") as out:
        for chunk in r.iter_content(chunk_size=1 << 20):
            out.write(chunk)

Relationship to Geo-trax

Songdo Vision is the detection dataset used to train and validate the default vehicle detector in Geo-trax, a pipeline that extracts georeferenced vehicle trajectories from high-altitude drone footage. The released YOLOv8s detector is published separately as the Hugging Face model rfonod/geo-trax.

The released Geo-trax detector was trained and validated on Songdo Vision v1, and the detection metrics on its model card refer to the v1 test split. It has not been retrained on v2.

This also matters for benchmarking: in v2, candidate missed vehicles were surfaced with a Geo-trax detector before human confirmation, so evaluating that same model family on the v2 test split carries a small favourable bias. It is bounded β€” 443 boxes, 0.8% of the test annotations β€” and filtering COCO id <= 272435 restores the v1-provenance test set for a strictly model-independent comparison.

Note: Songdo Vision annotates only the four vehicle classes above. The Geo-trax model was additionally trained on pedestrian and bicycle instances drawn from other sources, but those classes are not annotated here and are not recommended for use (see the model card for details).

The Geo-trax and Songdo family

All six Hub repos come from one multi-drone traffic experiment over Songdo, South Korea, and are gathered in the πŸ€— collection Trajectory Extraction and Stabilization.

Repo What it is
🧠 Model rfonod/geo-trax YOLOv8s vehicle detector for high-altitude drone imagery (0.951 mAP@50)
πŸš€ Space rfonod/geo-trax Run the detector in your browser on your own image or clip
πŸ›£οΈ Dataset rfonod/songdo-traffic ~700,000 georeferenced vehicle trajectories from 10 drones over 20 intersections
πŸ–ΌοΈ Dataset rfonod/songdo-vision (this page) 5,419 4K frames with 274,190 human-verified vehicle boxes
πŸ”„ Dataset rfonod/songdo-vision-obb The same frames with oriented (rotated) vehicle boxes
🎯 Dataset rfonod/stabilo-benchmark Benchmark data for reference-based UAV video stabilization

Related datasets and resources

Citation

If you use this dataset, please cite the associated publication:

@article{fonod2025advanced,
  title   = {Advanced computer vision for extracting georeferenced vehicle trajectories from drone imagery},
  author  = {Fonod, Robert and Cho, Haechan and Yeo, Hwasoo and Geroliminis, Nikolas},
  journal = {Transportation Research Part C: Emerging Technologies},
  volume  = {178},
  pages   = {105205},
  year    = {2025},
  doi     = {10.1016/j.trc.2025.105205}
}

Please also cite the dataset itself via its Zenodo record:

@dataset{fonod2026songdovision,
  author    = {Fonod, Robert and Cho, Haechan and Yeo, Hwasoo and Geroliminis, Nikolas},
  title     = {Songdo Vision: Vehicle Annotations from High-Altitude BeV Drone Imagery in a Smart City},
  year      = {2026},
  version   = {v2},
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.13828407},
  url       = {https://doi.org/10.5281/zenodo.13828407}
}

License

This dataset is released under the Creative Commons Attribution 4.0 International (CC BY 4.0) license; see the LICENSE file for the full terms. You are free to share and adapt the material for any purpose, provided you give appropriate credit.

Downloads last month
104

Models trained or fine-tuned on rfonod/songdo-vision

Space using rfonod/songdo-vision 1

Collection including rfonod/songdo-vision

Papers for rfonod/songdo-vision