Edge guided salient object detection

This paper proposes a novel edgtae guided salient object detection algorithm. Compared to existing works where edges artae always ignored, the proposed algorithm attaches great importance to edge information and shows that edges play an irreplaceable role throughout the salient object detection process. Specifically, we derive two scales of nested superpixels from an edge guided segmentation and then calculate a coarse saliency map on the finer scale by adopting color contrast, spatial prior and boundary prior. Next, a novel background prior is proposed by measuring geodesic distance among superpixels as the accumulated edge strength. Finally, coarse saliency, background prior, and inter-scale consistency are jointly integrated into an optimization function to obtain the final saliency map. Extensive experimental results on three benchmark datasets demonstrate that the proposed saliency model consistently outperforms the state-of-the-art saliency models.