UAV-GCD: UAV-based Ground Crack Dataset

We employed an unmanned aerial vehicle (UAV) to collect and construct a dedicated pavement crack dataset on highways in Anyang City, Henan Province, China. The dataset covers diverse real-world conditions, including sunny, cloudy, and shadow-occluded illuminations, as well as varying traffic states. Source videos were captured at an altitude of approximately 100 meters, with a resolution of 960 × 720 pixels and a duration of about 10 minutes per flight. Video frames were extracted at 25-frame intervals, yielding a total of 2,527 high-resolution aerial images. Pavement damage was subsequently annotated using the Labelme tool, resulting in 16,251 bounding box instances across six distinct distress categories (e.g., longitudinal cracks, alligator cracks, and potholes). To evaluate the model robustly, the dataset was rigorously partitioned into training, validation, and testing sets at an 8:1:1 ratio at the video level to prevent data leakage.