Action recognition using shared motion parts

Abstract

In this paper we analyse the advantages of a joint boosting method, previously known in the object recognition community, to detect and classify action keyframes. The method focuses on sharing object parts among action classes. Instead of sharing parts that only encode shape similarities, we propose to include motion information as an extra clue for sharing. We show that the inclusion of motion information significantly improves the recognition accuracy. The method is tested using a standard action database containing 10 action classes obtaining perfect classification. It also yields promising results on complicated videos including complex background.