Feature Decomposition and Attention-guided Boundary Refinement for Saliency Detection

In this paper, we propose a novel saliency detection model. We manage to balance the conflict between quality and efficiency by introducing Octave Convolution (OctConv). The proposed detection model is based on a typical encoder- decoder framework, and all encoded side-out features are de- composed into low and high frequency parts, respectively. The low frequency parts locate coarse salient regions, while the high frequency parts help capture edge details. And a two-stream attention module is proposed to emphasize important global and local context. A top-down attention transfers a global saliency prior to shallow layers and a bottom-up attention helps deep layers focus more on the edges. Compared to the state-of-the- art saliency methods, proposed model is of much lighter-weight (68.6 MB), runs faster (63.7 fps in real-time) and achieves competitive performances on six public benchmarks in terms of both quantitative and qualitative evaluation.