Unsupervised Image Co-Segmentation Based On Cooperative Game

In computer vision, co-segmentation is defined as the task of jointly segmenting the common objects in a given set of images. Most proposed co-segmentation algorithms have the assumptions that the common objects are singletons or with the similar size. In addition, they might assume that the background features are simple or discriminative. This paper presents a cooperative co-segmentation without these assumptions. In the proposed cooperative co-segmentation algorithm, each image is treated as a player. By using the cooperative game, heat diffusion, and image saliency, we design a constrained utility function for each player. This constrained utility function push all players, with the instinct to maximize their self-utility, to cooperatively define the common-object labels. We then use cooperative cut to segment the common objects according to the common-object labels. Experimental results demonstrate that the proposed method outperforms the state-of-the-art co-segmentation methods in the segmentation accuracy of the common objects in the images.