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

Overview The Songdo Vision dataset provides high-resolution 4K RGB images annotated with categorized axis-aligned bounding boxes (BBs) for vehicle detection from a high-altitude bird’s-eye view (BeV) perspective. Captured over Songdo International Business District, South Korea, it comprises 5,419 annotated video frames (4,335 training / 1,084 test; 80/20 split) containing 274,190 vehicle instances across four classes: Car (including vans and light-duty vehicles) Bus Truck Motorcycle This dataset can serve as a benchmark for aerial vehicle detection, supporting research and real-world applications in intelligent transportation systems, traffic monitoring, and aerial vision-based mobility analytics. It was developed in the context of a multi-drone experiment aimed at enhancing geo-referenced vehicle trajectory extraction [1]. A full dataset card with usage instructions and download guidance is available on Hugging Face. 📌 Citation: If you use this dataset in your work, kindly acknowledge it by citing the following article [1]: Robert Fonod, Haechan Cho, Hwasoo Yeo, Nikolas Geroliminis (2025). Advanced computer vision for extracting georeferenced vehicle trajectories from drone imagery, Transportation Research Part C: Emerging Technologies, vol. 178, 105205. DOI: 10.1016/j.trc.2025.105205. 🔗 Companion dataset: For precisely georeferenced vehicle trajectories extracted from the same large-scale multi-drone experiment, see Songdo Traffic: 10.5281/zenodo.13828384. Motivation Publicly available datasets for aerial vehicle detection often exhibit limitations such as: Non-BeV perspectives with varying angles and distortions Inconsistent annotation quality, with loose or missing bounding boxes Lower-resolution imagery, reducing detection accuracy, particularly for smaller vehicles Lack of annotation detail, especially for motorcycles in dense urban scenes with complex backgrounds To address these challenges, Songdo Vision provides high-quality human-annotated bounding boxes, with machine learning assistance used to enhance efficiency and consistency. This ensures accurate and reliable ground truth for training and evaluating detection models. Dataset Composition The dataset is randomly split into training (80%) and test (20%) subsets: Subset Images Car Bus Truck Motorcycle Total Vehicles Train 4,335 196,719 7,052 11,861 3,030 218,662 Test 1,084 49,862 1,766 3,080 820 55,528 Total 5,419 246,581 8,818 14,941 3,850 274,190 A subset of 5,274 frames was randomly sampled from drone video sequences, while an additional 145 frames were carefully selected to represent challenging cases, such as motorcycles at pedestrian crossings, in bicycle lanes, near traffic light poles, and around other distinctive road markers where they may blend into the urban environment. Data Collection The dataset was collected as part of a collaborative multi-drone experiment conducted by KAIST and EPFL in Songdo, South Korea, from October 4–7, 2022. A fleet of 10 drones monitored 20 busy intersections, executing advanced flight plans to optimize coverage. 4K (3840×2160) RGB video footage was recorded at 29.97 FPS from altitudes of 140–150 meters. Each drone flew 10 sessions per day, covering peak morning and afternoon periods. The experiment resulted in 12TB of 4K raw video data. More details on the experimental setup and data processing pipeline are available in [1]. Note on frame dimensions: 5,411 frames are 3840×2160 (landscape 4K). Eight frames: 05275–05281 in the training subset and 05282 in the test subset are 2160×3840 (portrait 4K), as recorded. Their annotations are correct for that orientation, and the per-image dimensions are stated explicitly in the COCO images entries and the Pascal VOC <size> element. Normalized YOLO coordinates are relative to each frame's own dimensions. Consumers that assume a fixed 3840×2160 frame size shall read the per-image dimensions instead. Bounding Box Annotations & Formats Annotations were generated using a semi-automated object detection annotation process in Azure ML Studio, leveraging machine learning-assisted bounding box detection with human verification to ensure precision. Each annotated frame includes categorized, axis-aligned bounding boxes, stored in three widely-used formats. All three are generated from a single canonical source, so they describe exactly the same boxes: the YOLO files carry sub-pixel precision, while the COCO and Pascal VOC files share one common integer rounding of those coordinates. COCO JSON format Single annotation file per dataset subset (i.e., one for training, one for testing). Contains metadata such as image dimensions, bounding box coordinates, and class labels. Example snippet: { "info": {"description": "Songdo Vision: …", "version": "2.0", "year": 2026}, "licenses": [{"id": 1, "name": "Creative Commons Attribution 4.0 International", "url": "https://creativecommons.org/licenses/by/4.0/"}], "images": [{"id": 1, "file_name": "00001.jpg", "width": 3840, "height": 2160}], "annotations": [ {"id": 1, "image_id": 1, "category_id": 2, "bbox": [500, 600, 200, 50], "area": 10000, "iscrowd": 0 } ], "categories": [ {"id": 1, "name": "car", "supercategory": "vehicle"}, {"id": 2, "name": "bus", "supercategory": "vehicle"}, {"id": 3, "name": "truck", "supercategory": "vehicle"}, {"id": 4, "name": "motorcycle", "supercategory": "vehicle"} ] } bbox is [x_min, y_min, width, height] in absolute pixels; area equals width × height. Image and annotation ids are globally unique across both subsets. Annotations carried over from v1 keep their original id (all ≤ 272435), so records can be traced between releases; annotations added in v2 are numbered from 1000001 upwards, which makes the two sets easy to tell apart. Ids are not contiguous. YOLO TXT format One annotation file per image, following the format: <class_id> <x_center> <y_center> <width> <height> Bounding box values are normalized to [0,1] relative to that frame's own dimensions, with the origin at the top-left corner. Every box lies fully inside the frame. Class ids follow names.txt: 0 = car, 1 = bus, 2 = truck, 3 = motorcycle. Example snippet: 0 0.520000 0.630000 0.100000 0.050000 2 0.250000 0.400000 0.150000 0.080000 Pascal VOC XML format One annotation file per image, structured in XML. Contains image properties and absolute pixel coordinates for each bounding box. <truncated> is set to 1 for boxes that touch a frame edge, 0 otherwise. Example snippet: <annotation> <folder>images</folder> <filename>00001.jpg</filename> <path>images/00001.jpg</path> <source><database>Songdo Vision</database></source> <size><width>3840</width><height>2160</height><depth>3</depth></size> <segmented>0</segmented> <object> <name>car</name> <pose>Unspecified</pose> <truncated>0</truncated> <difficult>0</difficult> <bndbox><xmin>500</xmin><ymin>600</ymin><xmax>600</xmax><ymax>650</ymax></bndbox> </object> </annotation> File Structure 1. Training Data (train.zip, ≈12.9 GB) train/ │── coco_annotations.json # COCO format │── images/ │ ├── 00001.jpg │ ├── … │── labels/ │ ├── 00001.txt # YOLO format │ ├── 00001.xml # Pascal VOC format │ ├── … 2. Testing Data (test.zip, ≈3.2 GB) test/ │── coco_annotations.json │── images/ │ ├── 00027.jpg │ ├── … │── labels/ │ ├── 00027.txt │ ├── 00027.xml │ ├── … Additional Files README.md – Dataset documentation (this description) LICENSE.txt – Creative Commons Attribution 4.0 License names.txt – Class names (one per line) data.yaml – Example YOLO configuration file for training/testing Changes in v2 v2 is an annotation-quality release. The images are byte-identical to v1 — only the annotation files, data.yaml and names.txt changed. Ids of surviving v1 annotations are preserved, so v1 and v2 records can be matched directly. In total 196 invalid boxes were removed and 1,951 missed vehicles were added, taking the dataset from 272,435 to 274,190 annotations: 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 Split across subsets, 157 removals and 1,508 additions fall in the training set, and 39 removals and 443 additions in the test set. Invalid annotations removed 196 of 272,435 boxes (0.072%) were removed. Each geometry threshold was fixed only after visually inspecting the boxes it discards, and every false_positive removal was inspected individually. Reason Removed What it catches false_positive 124 No vehicle at all: painted road markings, rooftop equipment, construction machinery, manhole covers side_too_small 43 Shorter side below 6 px (smallest genuine motorcycle side is ~12 px) duplicate 12 One vehicle annotated twice (IoU ≥ 0.5 and comparable box areas) extreme_aspect_ratio 9 Aspect ratio above 6 and shorter side below 35 px implausibly_large 5 Longer side above 500 px (largest genuine vehicle is ~403 px) curated 2 Single box spanning several diagonally parked cars area_too_small 1 Area below 150 px² (1st-percentile motorcycle area is ~520 px²) Total 196 The thresholds are deliberately far outside the legitimate distribution. In particular, the aspect-ratio rule also requires a short side below 35 px, because genuine articulated trucks in this dataset reach an aspect ratio of ~6.9 and would otherwise have been discarded. Detector-assisted review v1 was produced by a semi-automated annotation pass with human verification. That process is good at precision but can still miss objects, so v2 re-examined the annotations against an independent YOLOv8s vehicle detector trained on high-altitude BeV drone imagery (geo-trax). The detector was used only to rank candidates for human review — it never edited an annotation on its own, and every change below was confirmed by a human looking at the image. False positives removed. Every annotation was scored against the detector, and the boxes it did not support were inspected individually. Confirmed non-vehicles — road markings and painted text, rooftop plant, construction machinery, manhole covers — were removed. Missed vehicles added — 1,951 annotations. Confident detections (≥ 0.70) that overlapped no existing annotation were reviewed the same way; confirmed vehicles were added. This signal is stronger than it may appear: the detector was trained on these very annotations, so training actively pushed it to stay silent wherever the ground truth was empty. A confident detection in an unannotated spot is therefore good evidence of a genuine miss rather than a model artefact. Added boxes take their geometry and class from the detector, with a human confirming that a vehicle of that class is present. They carry COCO ids from 1000001 upwards, so anyone who wants only the original v1-provenance annotations can filter them out. Review coverage: all 1,983 addition candidates and all 1,053 false-positive candidates were inspected on contact sheets; every box marked for change was then re-checked zoomed. A random sample of 72 accepted additions was independently re-checked, of which 71 were correct, so a small residual error rate (order 1%) should be assumed for the added boxes. Consistency and correctness fixes All three formats now agree exactly. In v1, COCO truncated the box width while Pascal VOC truncated the box corner, so the two disagreed by up to 1 px on roughly 75% of boxes. Both are now derived from one shared rounding of the same source coordinates. 1,464 boxes extended microscopically past the frame border (by at most 0.002 px, a rounding artifact). All boxes are now strictly inside their frame, and every normalized YOLO coordinate lies within [0,1]. Duplicate annotations removed. Twelve vehicles carried two overlapping boxes, usually with conflicting class labels. A pair is treated as a duplicate only when the boxes overlap at IoU ≥ 0.5 and have comparable areas (ratio ≥ 0.5), so the axis-aligned box of a diagonally oriented bus can never absorb a small vehicle beside it. The tighter box was kept, except for one pickup truck where the larger box correctly covers both cab and bed. COCO gained the standard info and licenses blocks and a supercategory field on each category. Pascal VOC gained <path>, <pose>, <truncated> and <difficult> per object, which several standard VOC parsers require. data.yaml now points at the concept DOI (which always resolves to the latest version) instead of the v1-specific DOI, and its path field follows the Ultralytics convention: a relative path is resolved against the Ultralytics datasets directory, not against the location of data.yaml. The v1 value (../datasets/songdo_vision) only worked for one particular layout. names.txt now ends with a newline. The 8 portrait frames are documented (see Data Collection); they were undocumented in v1, which described every frame as 3840×2160. Known limitations Class boundaries. Car/van/truck assignments on v1 annotations were not revised. The detector disagreed with the annotated class on roughly 0.55% of boxes, but these are largely genuine judgement calls (e.g. large vans, pickups) that the original annotators owned. Remaining missed vehicles. The additions came from detections at confidence ≥ 0.70; vehicles the detector found with lower confidence, or missed entirely, are still unannotated. Recall is improved, not complete. Test-set provenance. Because the added test-split boxes were surfaced by a geo-trax detector, benchmarking that same model family on Songdo Vision test carries a small favourable bias. It is bounded — the additions are 0.8% of the test annotations, and every one was human-confirmed — but it is worth stating for strict benchmarking. Loosely overlapping pairs. 104 box pairs overlap at an IoU between 0.3 and 0.5, and one box fully contains two others. These are almost always correct — the axis-aligned box of a diagonally parked car naturally encloses its neighbours — so they were flagged rather than changed. Acknowledgments In addition to the funding sources listed in the metadata, the creators express their gratitude to Artem Vasilev for his dedicated efforts in data annotation. We also thank the research teams of Prof. Simon Oh (Korea University) and Prof. Minju Park (Hannam University) for their assistance during the data collection campaign, including the provision of drone equipment and student support. Citation & Attribution Preferred Citation: If you use Songdo Vision for any purpose — whether academic research, commercial applications, open-source projects, or benchmarking efforts — please cite our accompanying article [1]: Robert Fonod, Haechan Cho, Hwasoo Yeo, Nikolas Geroliminis (2025).Advanced computer vision for extracting georeferenced vehicle trajectories from drone imagery,Transportation Research Part C: Emerging Technologies, vol. 178, 105205.DOI: 10.1016/j.trc.2025.105205(Preprint: arXiv:2411.02136) BibTeX Note: Although Zenodo automatically provides a formal dataset citation (shown below), including citation export in various formats such as BibTeX, we kindly request that you reference the above article as the primary source of this work. Dataset Citation (for archival purposes): Robert Fonod, Haechan Cho, Hwasoo Yeo, Nikolas Geroliminis (2026). Songdo Vision: Vehicle Annotations from High-Altitude BeV Drone Imagery in a Smart City (v2). Zenodo. DOI: 10.5281/zenodo.13828407. The DOI above is the concept DOI, which always resolves to the latest version. To cite this specific release, use the version DOI shown on the Zenodo record page.