Image co-segmentation based on image complexity analysis and pixel-correspondences

Image co-segmentation refers to the task of segmenting jointly a set of images sharing objects of a specific class. Recent methods can co-segment simple images well, however their performance may degrade significantly on more cluttered images. In this paper a novel method is presented to jointly segment out common objects from a set of images using image complexity analysis and dense correspondences. The images are first ranked based of their complexity scores. Then a set of simple images is selected and segmented. Finally the segmentation results of simple images are propagated to more cluttered images via multi-scale dense correspondences in order to guide their segmentation. The comparison experiments conducted on iCoseg dataset demonstrate the performance of the proposed method.