A Shadow Detection Method for Retaining Key Objects in Complex Scenes

The existing shadow detection methods have achieved good results on standard shadow datasets such as SBU and UCF. However, in actual large-scale scenes, key objects covered by shadows are often regarded as shadows, which may harm computer vision tasks. In the paper, we are the first to propose the Object-aware Shadow Detection Network (OSD-Net) model for computer vision tasks in complex scenes. It introduces the direction-aware spatial context (DSC) module to detect shadows, uses semantic segmentation with Mask RCNN to extract key objects in the picture, and designs a function to perform mask fusion. Qualitative experiments have been performed to test OSD-Net on three public datasets commonly used in computer vision. Compared with popular shadow detection methods, OSD-Net is able to effectively protect the key targets in the picture from being misjudged as shadows, and ensure shadow detection accuracy.