Don’t worry about noisy labels in soft shadow detection

Soft shadow is harder to detect than hard shadow as its complex characteristics (i.e., low-contrast, irregular shape, and ambiguous shadow boundaries). To improve the detecting capacity of these images, in this paper, we create a new benchmark for soft shadow detection and then design a reasonable supervision strategy to alleviate the effect of annotation noises. Next, we present a general shadow detection framework based on transformer to deal with complex scenes. Concretely, we combine the traditional channel attention and recent popular self-attention into our network. Moreover, we introduce a deep supervision mechanism that performs deep layer supervision to “guide” early classification results at each layer, which can further improve our detection performance. Finally, experimental results on three datasets show that our shadow transformer can be favorable against current state-of-the-art detectors.