In this paper, a discriminative two-phase dictionary learning framework isproposed for classifying human action by sparse shape representations, in whichthe first-phase dictionary is learned on the selected discriminative frames andthe second-phase dictionary is built for recognition using reconstructionerrors of the first-phase dictionary as input features. We propose a "zerothclass" trick for detecting undiscriminating frames of the test video andeliminating them before voting on the action categories. Experimental resultson benchmarks demonstrate the effectiveness of our method.
展开▼