rfonod/songdo-vision-obb

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

Songdo Vision OBB is the oriented bounding box counterpart of Songdo Vision. It provides the same 5,419 4K bird's-eye view (BEV) drone frames over the Songdo International Business District, South Korea, with every one of the 274,190 vehicle annotations expressed as a rotated rectangle instead of an axis-aligned box, across the same four classes.

Horizontal boxes are a poor fit for high-altitude BEV imagery, where vehicles appear at arbitrary in-plane orientations: a diagonally traveling or parked vehicle fills a small fraction of its axis-aligned box and the rest is road surface, lane markings, or a neighboring vehicle. This release replaces those boxes with tight oriented ones, so the dataset can serve as a benchmark for oriented aerial vehicle detection and as a source of vehicle heading for downstream traffic analysis.

📦 The data is hosted on Zenodo, not on Hugging Face. This page is a documentation mirror and index card. The full dataset (images + annotations, ≈16.4 GB) lives immutably on Zenodo; this card describes v1, DOI 10.5281/zenodo.15050578. See Access the dataset below for how to download it.

Dataset at a glance

Property	Value
Modality	4K RGB images, bird's-eye view (BEV), 3840 × 2160, except 8 frames recorded portrait (2160 × 3840)
Frames	5,419 (4,335 train / 1,084 test; 80/20 split, identical to Songdo Vision)
Instances	274,190 (218,662 train / 55,528 test), one per horizontal box in Songdo Vision v2
Classes	4: 0 car (incl. vans), 1 bus, 2 truck, 3 motorcycle
Annotation type	Oriented (rotated) bounding boxes, four corners, clockwise in image coordinates
Formats	YOLO OBB TXT (normalized, canonical), DOTA TXT, COCO JSON
Per-object side-cars	conversion confidence, and the segmentation contour each box was fitted to
Tuning and validation set	50 additional frames with 4,245 manually annotated oriented boxes (every rotated vehicle redrawn by hand), plus the full grid search
Acquisition	Fleet of 10 drones over 20 intersections, 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	≈16.4 GB (train.zip + test.zip + hbb2obb_benchmark.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

How the oriented boxes were produced

The boxes were converted, not re-annotated. Producing 274k oriented boxes by hand is not practical, and the vehicle inventory and class labels of Songdo Vision v2 are already human-verified, so that work carries over intact.

Each horizontal box is used as a prompt for a SAM segmentation model, the mask is clipped back to the prompt region and cleaned, and the minimum-area rotated rectangle of the resulting contour becomes the oriented box. Every box inherits the class label of the horizontal box it came from; nothing is re-classified and no annotation is dropped, so the instance count matches Songdo Vision v2 exactly. The tool is HBB2OBB, written for this purpose and released separately; this release was converted by HBB2OBB v1.10.0 (10.5281/zenodo.22664151), the version each subset's PROVENANCE.txt records.

The four hyperparameters were selected by grid search (11 model sets, 495 grid points) against the 50 manually annotated frames shipped with the release, which share no image with Songdo Vision. The settings behind this release are sam_l at inference size 1024, prompt scale factor 0.02, morphological kernel -0.1 (closing), fragment ratio 0.1.

Measured on the 50 manual frames	Value
Mean / median IoU against the manual boxes	0.88679 / 0.91759
Share of boxes at IoU ≥ 0.75 / 0.90	92.06% / 61.51%
Orientation error, 90th percentile	3.50°
Boxes that fell back to their horizontal box	0 of 4,245

Those 50 frames were used to select the settings, so the figures characterize the method rather than benchmark it on held-out data. The grid search, the figures behind it, the manual boxes themselves and the scripts behind the confidence and prompt-robustness tables all ship in hbb2obb_benchmark.zip, so the selection can be repeated or a different method measured on the same footing.

Orientation is derived, so check it before you trust it

No human inspected the oriented geometry of the 5,419 train and test frames. Two things follow.

Use the confidence side-car. labels_confidence/<frame>.txt holds one score in [0, 1] per line, row-aligned with labels/<frame>.txt, measuring how well the segmentation fit its box. It is the single most useful thing in this release if orientation accuracy matters to you: on the validation frames it tracks the IoU actually achieved, and a cut at 0.80 removes 58% of the boxes below IoU 0.75. That threshold is also what sets the difficult flag in the DOTA files, so standard DOTA protocols can exclude them without reading the side-car at all. Across the release the scores run from 0.4025 to 1.0, with a median of 0.9154, and the 0.80 cut flags 14,404 boxes (5.25%).

A fallback would be marked 0.0. Where segmentation finds no usable mask inside a box, the original axis-aligned box is kept as the oriented one, which is invisible in the label file alone. No box in this release is one: a usable mask was found inside all 274,190 horizontal boxes. A low score still means a poor fit, so filter on the score rather than on this.

Motorcycles convert poorly (mean IoU 0.785 against 0.890 for cars, 90th-percentile orientation error 13.9° against 3.3°): a motorcycle is about 18 px across its shorter side, and rider and machine do not always segment as one object. Near-square vehicles are also inherently ambiguous, since the minimum-area rectangle can rotate far under a one-pixel change in the mask.

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:

➡️ v1: DOI 10.5281/zenodo.15050578 (the version DOI of this release; the record page also shows the concept DOI, which always resolves to the latest version)

Files on Zenodo:

File	Size	Contents
train.zip	≈12.98 GB	Training images + oriented annotations (YOLO OBB / DOTA / COCO) + both side-cars
test.zip	≈3.24 GB	Test images + oriented annotations + both side-cars
hbb2obb_benchmark.zip	≈169 MB	The 50 manually annotated frames, the complete tuning study and the scripts behind its tables
names.txt	small	Class names
data.yaml	small	Ready-to-use Ultralytics OBB dataset configuration
README.md	small	Dataset documentation
LICENSE.txt	small	CC BY 4.0 terms
SHA256SUMS	small	SHA-256 of every other file (sha256sum -c SHA256SUMS)

Download with zenodo_get:

pip install zenodo_get
zenodo_get 10.5281/zenodo.15050578      # fetches all files of v1

Or directly with requests:

import requests

# v1. For whichever version is the latest, follow record["conceptdoi"] instead.
record = requests.get("https://zenodo.org/api/records/15050578").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)

Train on it straight away. The YOLO .txt files are already in the normalized [0, 1] convention Ultralytics reads for the OBB task, and data.yaml sets no path:, so unzipping the two archives beside it is the whole setup:

yolo obb train data=data.yaml model=yolo26n-obb.pt imgsz=1280

The same boxes in absolute pixels are in the .dota files, for MMRotate and the DOTA devkit, and in the COCO files, with the quad in segmentation and its axis-aligned envelope in bbox.

Relationship to Songdo Vision

Songdo Vision OBB is a derivative of Songdo Vision v2, produced under its CC BY 4.0 terms. It is not an independent annotation effort: the imagery is byte-identical, and the vehicle inventory, the class assignments and the annotation-quality characteristics are those of Songdo Vision v2, including its known limitations. What is new is the geometry.

The horizontal boxes are not duplicated here. They stay in Songdo Vision, where all three of their formats are maintained; a copy under a second DOI would fork them. The dataset README gives the two unzip lines that bring them back for anyone repeating the conversion.

COCO annotation and image ids are preserved, so a box in this dataset and its source horizontal box join directly by id. Annotations that originated in Songdo Vision v1 keep ids ≤ 272435; those added in v2 carry ids from 1000001 upwards. Image ids are the global frame number, so the train and test files share none.

Related datasets and resources

• Songdo Vision: the source horizontal-box dataset, v2 being the version converted: 10.5281/zenodo.21728430 (v2; concept DOI 10.5281/zenodo.13828407) · HF rfonod/songdo-vision
• Songdo Traffic: companion georeferenced trajectory dataset (latest version): 10.5281/zenodo.13828383 · HF rfonod/songdo-traffic
• HBB2OBB: the conversion and evaluation tool that produced this dataset: github.com/rfonod/hbb2obb · v1.10.0, the version that converted it, 10.5281/zenodo.22664151 (concept DOI 10.5281/zenodo.15151143)
• stabilo-benchmark: the benchmark of the paper that introduces this dataset, which annotates its own data at the settings selected here: github.com/rfonod/stabilo-benchmark · v1.0.0, 10.5281/zenodo.24168256 · its data record 10.5281/zenodo.23623540
• Geo-trax detector: HF model rfonod/geo-trax · github.com/rfonod/geo-trax
• Publications: the paper introducing this dataset (arXiv:2611.17190), and the article behind the imagery, Transportation Research Part C (2025): 10.1016/j.trc.2025.105205 · arXiv:2411.02136

Citation

If you use this dataset, please cite the paper that introduces it, which describes the conversion and its validation (arXiv:2611.17190):

@article{fonod2026benchmark,
  author  = {Fonod, Robert and Geroliminis, Nikolas},
  title   = {Consensus Capture in Reference-Based UAV Video Stabilization: A Benchmark on Dynamic Urban Scenes},
  journal = {arXiv preprint arXiv:2611.17190},
  year    = {2026},
  doi     = {10.48550/arXiv.2611.17190}
}

Please also cite the article behind the imagery and the source annotations:

@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}
}

And the dataset itself, via its Zenodo record (v1):

@dataset{songdo_vision_obb,
  author    = {Fonod, Robert and Geroliminis, Nikolas},
  title     = {Songdo Vision OBB: Oriented Bounding-Box Vehicle Annotations from High-Altitude BeV Drone Imagery in a Smart City},
  year      = {2026},
  version   = {v1},
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.15050578},
  url       = {https://doi.org/10.5281/zenodo.15050578}
}

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.