DAB-Sky Multitemporal Drone Dataset, 2025

The DAB-Sky dataset comprises 39,827 multitemporal aerial images that integrate both real and synthetic drone imagery for use in drone detection, segmentation, and airspace monitoring research. Real-world images were collected using DSLR and 2K industrial surveillance cameras positioned at varying distances and altitudes. To capture illumination diversity, image acquisition was conducted during four distinct time periods: morning, noon, evening, and night. Complementing the real imagery, synthetic images were produced using Blender 3D and Unreal Engine, where drone models were rendered under adjustable lighting conditions, sky environments, and atmospheric scattering. Aerodynamic variations—including pitch shifts of ±25°, roll adjustments of ±20°, and yaw movements of ±30° were incorporated to simulate realistic flight dynamics, alongside additional visual effects such as motion blur, exposure variation, and HDRI-based illumination.
Every image in the dataset is accompanied by two forms of annotations: YOLO-format bounding boxes for object detection tasks and polygon segmentation masks for fine-grained contour analysis. All images were originally recorded at a resolution of 1920×1080 pixels and subsequently standardized to 450×450 pixels to ensure uniformity across the dataset. Through its combination of diverse lighting conditions, varied backgrounds, multiple drone orientations, and dual-domain (real and synthetic) representation, DAB-Sky provides a comprehensive resource suitable for benchmarking drone detection methods, training computer vision models, and supporting research in intelligent airspace surveillance.