An automatic method for detecting objects of interest in videos using surprise theory

Automatically detecting objects of interest in videos is a challenging issue since there is no prior knowledge about which objects should be detected and what these objects look like. The objects of interest can be defined as salient ones and the saliency can be measured by surprise theory. Therefore, this paper proposes a new method for automatic object detection. It involves two modules: surprise estimation and object localization. The surprise estimation module first uses the surprise theory to obtain a saliency map which indicates the novelty of each pixel compared with its previous states. The object localization module then determines where the salient objects locate based on the branch-and-bound search algorithm. Experimental results have shown that the objects of interest in videos can be successfully localized by using the proposed automatic detection method.