Saliency detection with two-level fully convolutional networks

This paper proposes a deep architecture for saliency detection by fusing pixel-level and superpixel-level predictions. Different from the previous methods that either make dense pixellevel prediction with complex networks or region-level prediction for each region with fully-connected layers, this paper investigates an elegant route to make two-level predictions based on a same simple fully convolutional network via seamless transformation. In the transformation module, we integrate the low level features to model the similarities between pixels and superpixels as well as superpixels and superpixels. The pixel-level saliency map detects and highlights the salient object well and the superpixel-level saliency map preserves sharp boundary in a complementary way. A shallow fusion net is applied to learn to fuse the two saliency maps, followed by a CRF post-refinement module. Experiments on four benchmark data sets demonstrate that our method performs favorably against the state-of-art methods.