A Fast and Efficient Network for Single Image Shadow Detection

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