UAV-OBB: An Aerial Urban Vehicle Dataset with Oriented Bounding Boxes for Remote Sensing Object Detection in Smart Cities
Description
UAV-OBB is an aerial urban vehicle dataset designed for rotation-aware object detection and smart-city traffic monitoring using oriented bounding boxes (OBBs). Many UAV/drone datasets annotate vehicles with axis-aligned rectangles that include unnecessary background and do not encode orientation; UAV-OBB provides tight rotated annotations to support accurate localization, orientation estimation, and downstream traffic analysis. The dataset contains 1,617 RGB images (JPEG) at 1920×1080 resolution captured from predominantly nadir-view UAV imagery over urban roads in Chongqing, and Wuhan. It provides 46,807 oriented vehicle instances across six classes: bike, bus, car, other_vehicle, taxi, and truck. The data are split into 1,383 training images, 218 validation images, and 16 test images. Collection altitude was approximately 75-108 m under diverse real-world conditions, including morning, midday, evening, rain, night, and mist/light fog. Both wide field-of-view and zoom settings were used to introduce strong scale variation, ranging from small distant vehicles to large close-up buses and trucks. Annotations are provided in YOLOv8-OBB label format as plain text files (one per image). Each object is encoded by class id and a rotated rectangle represented by four corner vertices (x1,y1…x4,y4) in normalized image coordinates. All instances were manually annotated with rotation-capable tools and double-checked for quality and consistency. Occluded and truncated vehicles were included when the majority of the object was visible, reflecting realistic urban traffic scenes. Directory structure (YOLO-style) The dataset is released in a YOLO-style directory layout. The root folder UAV-OBB/ contains the train/validation/test splits, the dataset configuration file (data.yaml), and supplementary MP4 evaluation videos, organized as follows: UAV-OBB/ ├─ train/ │ ├─ images/ │ └─ labels/ ├─ valid/ │ ├─ images/ │ └─ labels/ ├─ test/ │ ├─ images/ │ └─ labels/ ├─ data.yaml └─ test_videos_mp4/ All images are 1920 × 1080 pixels (8-bit RGB) in resolution and in JPEG format, captured by UAV cameras in two Chinese cities. To support practical evaluation beyond static images, UAV-OBB also includes supplementary MP4 video sequences: a short clip with sparsely annotated reference frames for human-verifiable temporal assessment, and two longer unannotated sequences for qualitative evaluation of stability and deployment behavior. UAV-OBB can be used for oriented detection, tracking, counting, density estimation, and traffic-flow analysis in urban UAV surveillance scenarios. If you use this dataset in your publication or thesis, please cite: Ahmad, Israr; Fengjun, Shang; Bibi, Kiran; Slaman Pathan, Muhammad (2026), “UAV-OBB: An Aerial Urban Vehicle Dataset with Oriented Bounding Boxes for Remote Sensing Object Detection in Smart Cities”, Mendeley Data, V3, doi: 10.17632/6snrjwcpkh.4
Files
Institutions
- Chongqing University of Posts and TelecommunicationsChongqing, Chongqing