K-Vehicles: A Remote Sensing Dataset for Vehicle Detection in Aerial Imagery

Announcements

• K-Vehicles is now available for download at Hugging Face
• K-Vehicles has been accepted at the Second IEEE/CVF Workshop on Computer Vision for Geospatial Image Analysis (GeoCV) @ WACV 2026 📣📣📣
• K-Vehicles is under review (download option will be available after publication)

About the project

We present K-Vehicles, a new dataset for vehicle detection in aerial imagery. It is built from high-resolution RGB images captured by a Cessna aircraft over diverse real-world environments, including highways, agricultural fields, and industrial zones. The dataset comprises 15,168 cropped images of 1,024x1,024 pixels, annotated manually across seven vehicle categories: truck, forklift, machinery, pickup, tractor, car, and bus. It incorporates relevant challenges such as occlusion, scene clutter, intra-scene variation, and variable lighting conditions, making it suitable for training and evaluating object detection models in realistic scenarios.

Paper

K-Vehicles paper is available at WACV proceedings here

Data set description

K-Vehicles dataset consists of 15,168 cropped images, each with a fixed resolution of 1024x1024 pixels. These patches were obtained from aerial images captured under diverse conditions. The dataset is split into training, validation, and test subsets following an 80-10-10 proportion, resulting in 12,134 images for training, and 1,517 images for both validation and test sets. Annotations were generated in YOLO format, with each object instance described by its class and normalized bounding box coordinates. In total, the dataset includes 63,233 annotated instances, distributed as follows: 50,993 in the training set, 5,961 in the validation set, and 6,279 in the test set. The summary of K-Vehicles is as follows:

Split	# Images
Train	12,134
Valid	1,517
Test	1,517

Dataset download

The dataset is available for download via Hugging Face here.

Benchmarking results

To assess the utility of K-Vehicles, we conduct a benchmarking study using four recent YOLO architectures, from YOLOv9 to YOLOv12.

Model	Precision	Recall	mAP@50	mAP@50:95	Speed
YOLOv9m	0.956	0.770	0.865	0.698	2.9
YOLOv10m	0.910	0.788	0.844	0.687	2.0
YOLOv11m	0.960	0.769	0.865	0.700	2.4
YOLOv12m	0.931	0.782	0.883	0.736	3.5

Contributing

Contributions are what make the open source community such an amazing place to learn, inspire, and create. Any contributions you make are greatly appreciated.

If you have a suggestion that would make this better, please fork the repo and create a pull request. You can also simply open an issue with the tag "enhancement". Don't forget to give the project a star!

1. Fork the Project
2. Create your Feature Branch (git checkout -b feature/NewFeature)
3. Commit your Changes (git commit -m 'Add some NewFeature')
4. Push to the Branch (git push origin feature/NewFeature)
5. Open a Pull Request

License

Distributed under GNU General Public License v3.0. See LICENSE for more information.

BibTex

If you find this dataset useful, please star ⭐️⭐️⭐️ our repo and cite our paper.

@InProceedings{Ramos_2026_WACV,
    author    = {Ramos, Leo Thomas and Casas, Edmundo and Romero, Cristian and Rivas-Echeverr{\'\i}a, Francklin},
    title     = {K-Vehicles: A Remote Sensing Dataset for Vehicle Detection in Aerial Imagery},
    booktitle = {Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) Workshops},
    month     = {March},
    year      = {2026},
    pages     = {800-809}
}

Poster