Abstract

With the goal of effectively identifying common and  salient objects in a group of relevant images, co-saliency  detection has become essential for many applications such  as video foreground extraction, surveillance, image  retrieval, and image annotation. In this paper, we propose  a unified co-saliency detection framework by introducing  two novel insights: 1) looking deep to transfer higher-level  representations by using the convolutional neural network  with additional adaptive layers could better reflect the  properties of the co-salient objects, especially their  consistency among the image group; 2) looking wide to  take advantage of the visually similar neighbors beyond a  certain image group could effectively suppress the  influence of the common background regions when  formulating the intra-group consistency. In the proposed  framework, the wide and deep information are explored for  the object proposal windows extracted in each image, and  the co-saliency scores are calculated by integrating the  intra-image contrast and intra-group consistency via a  principled Bayesian formulation. Finally the window-level  co-saliency scores are converted to the superpixel-level  co-saliency maps through a foreground region agreement  strategy. Comprehensive experiments on two benchmark  datasets have demonstrated the consistent performance  gain of the proposed approach.1