Multi-Background Airborne Object Detection (MBAOD)

The Multi-Background Airborne Object Detection dataset (MBAOD) is a curated benchmark for detecting airborne objects under diverse environmental conditions. It contains six object classes: airplane, bird, helicopter, kite, parachute, and UAV. The dataset includes objects at different scales, orientations, and positions across varied backgrounds, including sky, field, forest, desert, mountain, lake, urban, and semi-urban scenes. It contains background-only, single-object, and multi-object images, with an average of approximately 1.6 annotated objects per image. MBAOD was compiled from multiple sources. Video sequences were converted into individual frames, while image samples were incorporated directly. After manual curation to remove low-quality, duplicate, redundant, or irrelevant samples, the data were organized, annotated, and reviewed using the Roboflow platform. Annotations are provided in TXT, JSON, XML, and CSV formats for compatibility with different experimental pipelines. The dataset is released in two main versions: C-MBAOD, containing 35,106 clean images and 57,000 balanced bounding-box annotations, and WA-MBAOD, containing 59,612 images and 97,092 annotations, including additional synthetically generated adverse-weather training samples. Four dedicated weather test sets—Dusty (DTS), Foggy (FTS), Rainy (RTS), and Snowy (STS)—are also provided for controlled robustness evaluation under visually degraded conditions. MBAOD is intended to support research in airborne object detection and tracking, particularly for UAVs, with potential applications in aerial surveillance and security, airspace monitoring and management, search-and-rescue support, critical infrastructure protection, and related applications.