VRules: an effective association-based classifier for videos

Abstract

Video classification is an important step towards multimedia understanding. Most state-of-the-art approaches which apply HMM to capture the temporal information of videos have the limitation by assuming that the current state of a video depends only on the immediate previous state. Nevertheless, this assumption may not hold for videos of various categories. In this paper, we present an effective video classifier which employs the association rule mining technique to discover the actual dependence relationship between video states. The discriminatory state transition patterns mined from different video categories are then used to perform classification. Besides capturing the association between states in the time space, we also capture the association between low-level features in spatial dimension to further distinguish the semantics of videos. Experimental results show that the performance of our association rule based classifier is quite promising.