Abstract
Co-segmentation is widely discussed in recent years. The main goal of co-segmentation is simultaneously segmenting the same or similar objects from a set of images. Unlike previous methods, in this paper, we think about this problem from different angles. Due to the central part of co-segmentation is to find the common regions among images, we propose an extensive-form game based matching algorithm to solve this problem. In the proposed matching game, each image represents a rational player, and their strategies are their superpixels. Players purpose to find some matched pairs with higher matching score in order to get higher payoff. In our method, according to the color and texture feature, we utilize the proposed matching game to find matched pairs with higher matching score. Here we assume that the matched pairs between common regions usually have higher score than others. So we can locate the common foreground in each image by choosing a proper threshold. Then we apply graph cut to segment the foreground regions of each images independently. We show the superior performance of our method in comparison with other state-of-the-art techniques on some realistic dataset.