On Multiple Image Group Cosegmentation

The existing cosegmentation methods use intra-group information to extract a common object from a single image group. Observing that in many practical scenarios there often exist multiple image groups with distinct characteristics but related to the same common object, in this paper we propose a multi-group image cosegmentation framework, which not only discoveries intra-group information within each image group, but also transfers the inter-group information among different groups so as to more accurate object priors. Particularly, we formulate the multi-group cosegmentation task as an energy minimization problem. Markov random field (MRF) segmentation model and dense correspondence model are used in the model design and the Expectation-Maximization algorithm (EM) is adapted to solve the optimization. The proposed framework is applied on three practical scenarios including image complexity based cosegmentation, multiple training group cosegmentation and multiple noise image group cosegmentation. Experimental results on four benchmark datasets show that the proposed multi-group image cosegmentation framework is able to discover more accurate object priors and significantly outperform state-of-the-art single-group image cosegmentation methods.