RMLANet: Random Multi-Level Attention Network for Shadow Detection

Abstract

This paper addresses the problem of shadow detection from a single image. Previous approaches have shown that exploiting both global and local contexts in deep convolutional neural network layers can greatly improve performance. However, multi-level contexts remain underexplored. To achieve this, we propose RMLANet, a novel Random Multi-Level Attention Network. Specifically, we leverage shuffled multi-level features simultaneously with guiding features, and employ the transformer to capture global context. Furthermore, to reduce the computational and memory overhead caused by the self-attention mechanism in the vanilla transformer, we propose a random sampling strategy to reduce the number of inputs to the transformer. This is motivated by observing local consistency in images, which suggests that dense attention is unnecessary. Extensive experimental results demonstrate that our method outperforms current state-of-the-art methods on three widely used benchmark datasets SBU, ISTD and UCF.