Airborne-Shadow: Towards Fine-Grained Shadow Detection in Aerial Imagery

Abstract

Shadow detection is the first step in the process of shadow removal, which improves the understanding of complex urban scenes in aerial imagery for applications such as autonomous driving, infrastructure monitoring, and mapping. However, the limited annotation in existing datasets hinders the effectiveness of semantic segmentation and the ability of shadow removal algorithms to meet the fine-grained requirements of real-world applications. To address this problem, we present Airborne-Shadow (ASD), a meticulously annotated dataset for shadow detection in aerial imagery. Unlike existing datasets, ASD includes annotations for both heavy and light shadows, covering various structures ranging from buildings and bridges to smaller details such as poles and fences. Therefore, we define shadow detection tasks for multi-class, single class, and merging two classes. Extensive experiments show the challenges that state-of-the-art semantic segmentation and shadow detection algorithms face in handling different shadow sizes, scales, and fine details, while still achieving comparable results to conventional methods. We make the ASD dataset publicly available to encourage progress in shadow detection.