Saliency detection or foreground segmentation is a fundamental and important task in computer vision, which can be treated as a pixel-wise classification problem. Recently, although fully convolutional network (FCN) based approaches have made remarkable progress in this task, segmenting salient objects in complex image scenes is still a challenging problem. In this paper, we argue that, when predicting the saliency of a given pixel, human-like attention mechanisms play an important role in structural saliency inference. Therefore, we propose a simple yet surprisingly effective self-gated soft-attention mechanism for fast saliency detection. The soft-attention mechanism generates a gating signal that is end-to-end trainable, which allows deep networks to contextualize local information useful for saliency prediction. In addition, the proposed attention mechanism is channel-wise, generic and can be easily incorporated into any existing FCN architectures like Trojan Horse, while only requiring negligible parameters. Extensive experiments verify the superior effectiveness of the proposed method. More specifically, our method achieves a new state-of-the-art performance on seven public saliency benchmarks, and outperforms the very recent methods with a large margin.
View on arXiv