A site entropy rate and degree centrality based algorithm for image co-segmentation

Abstract

In this paper, we propose a graph based algorithm that efficiently segments common objects from multiple images. We first generate a number of object proposals from each image. Then, an undirected graph is constructed based on proposal similarities and co-saliency maps. Two different methods are followed to extract the proposals containing common objects. They are: (1) degree centrality of nodes obtained after graph thresholding and (2) site entropy rate of nodes calculated on the stationary distribution of Markov chain constructed on the graph. Finally, we obtain the co-segmented image region by selecting the more salient of the object proposals obtained by the two methods, for each image. Multiple instances of the common object are also segmented efficiently. The proposed method has been compared with many existing co-segmentation methods on three standard co-segmentation datasets. Experimental results show its effectiveness in co-segmentation, with larger IoU values as compared to other co-segmentation methods.