A FAST AND EFFICIENT NETWORK FOR SINGLE IMAGE SHADOW DETECTION

Shadows in image degrades performance of many applications. In this paper, we propose a novel multi-level featureaware network, called TransShadow, which uses transformer to capture both the local context and global context for shadow detection from a single image. Specifically, we design a multi-level feature-aware module, where several levels of features are selected and processed by transformer to distinguish the shadow and non-shadow areas. To further utilize the remaining levels of features, progressive upsampling with skip connections are proposed for fusing more information for shadow detection. Experimental results show that our approach achieves comparative performance as the state-of-the-art method on benchmark dataset SBU and ISTD with the smallest model size and the fastest inference speed. More importantly, our model shows the best generalization performance on benchmark dataset UCF.