Graph based spatiotemporal saliency detection incorporating low and high level features

In this paper, we propose a novel graph based spatiotemporal saliency detection method which models eye movements using random walk with restart. The method is executed in superpixel domain and unify low level features and high level features into the framework of random walk with restart. The boundary prior, as a high level feature, is employed to obtain a boundary prior based restarting distribution. The temporal saliency map, which is achieved utilizing the low level motion features, is regarded as another restarting distribution. Then the spatiotemporal saliency map is implemented by incorporating two restarting distributions and spatial transition matrix into the random walk with restart framework. Experiment results tested on two public databases show that the proposed method outperforms the existing saliency detection methods.