A Hybrid Graph Model for Unsupervised Object Segmentation

Abstract

In this work, we address the problem of performing class specific unsupervised object segmentation, i.e., automatic segmentation without annotated training images. We propose a hybrid graph model (HGM) to integrate recognition and segmentation into a unified process. The vertices of a hybrid graph represent the entities associated to the object class or local image features. The vertices are connected by directed edges and/or undirected ones, which represent the dependence between the shape priors of the class (for recognition) and the similarity between the color/texture priors within an image (for segmentation), respectively. By simultaneously considering the Markov chain formed by the directed subgraph and the minimal cut of the undirected subgraph, the likelihood that the vertices belong to the underlying class can be computed. Given a set of images each containing objects of the same class, our HGM based method automatically identifies in each image the area that the objects occupy. Experiments on 14 sets of images show promising results.