"Beach-Litter-UAV dataset"

"To address the critical lack of benchmark datasets tailored for extremely small beach-litter object (ESO) detection in UAV imagery, we construct the Beach-Litter-UAV dataset by systematically screening two publicly available UAV garbage datasets and integrating them with a newly collected Zhuhai Coastal Beach Litter Dataset acquired along the coast of Zhuhai City, China. The dataset comprises high-resolution RGB images captured under diverse coastal conditions, including sandy beaches, rocky shorelines, and vegetated areas, with centimeter-level ground sampling distances. A distinctive characteristic of Beach-Litter-UAV is its focus on ESOs: over 80% of the 6,338 annotated instances across five common litter categories (plastic, flexible_pouch, metal, glass, and food_container) have normalized sizes below 0.02, meaning they occupy only a few tens of pixels even in high-resolution imagery. This extreme scale, combined with complex background textures such as sand ripples, shells, and wet\u2013dry transitions, makes the dataset uniquely challenging for existing object detectors and ideal for developing and evaluating algorithms specialized in ESO recognition. By providing a standardized benchmark with rich annotations and realistic coastal scenarios, Beach-Litter-UAV aims to foster progress in automated beach-litter monitoring and support environmental protection efforts along global shorelines."