Matching based image co-segmentation

The goal of image co-segmentation is to segment the same or similar objects from a set of images. Unlike traditional methods, we propose a matching based algorithm to achieve this goal. Our method contains two phases. In the first phase, we use a matching algorithm to jointly estimate initial foreground labels of the input images. In the next phase, the labels of each image are used to extract its foreground regions via graph cuts. In contrast to other co-segmentation algorithms, our approach decomposes the co-segmentation problem into the two simpler phases, thus preventing the need to construct a complicated co-segmentation graph model which may cause troublesome optimization. The experimental results show the competitive performance of the proposed method in comparison with other famous image co-segmentation techniques on the CMU-Cornell iCoseg dataset that has variability in object deformations and poses.