Exemplar Based Image Salient Object Detection

Saliency detection is an important problem. Researchers in this area mainly focus on advanced models to achieve high performance on benchmark datasets with a large number of labeled images. However, most conventional saliency detection methods only use these benchmark datasets for saliency evaluation. We argue that we can use these valuable labeled data to generate precise saliency results. In this paper, we propose to exploit these labeled data by retrieving labeled images with similar foreground to a query image as exemplars. Then we learn to generate precise saliency from these exemplars. We conduct extensive experiments on four benchmark datasets, and we compare our method with eleven stateof-the-arts. Experimental results show the promising performance improvements of our method over compared methods.