Human action recognition from videos draws tremendous interest in the pastmany years. In this work, we focus on the task of action recognition that can be statedas: Given a video sequence obtaining one person performing an action, the task is todesign a system for automatically recognizing what the action is performed in thisvideo.In this work, we perform the task of action recognition under three views that arecomprised of two template views and one testing view. We first find that the trifocaltensor resides in a twelve dimensional subspace of the original space if he first twoviews are already matched and the fundamental matrix between them is known, whichwe refer to as subtensor. And the trifocal tensor represented by the subtensor satisfiesall the internal constraints. Then we use the subtensor to perform the task of actionrecognition under three views.We find that:1) Recognizing actions under three views is more effective thantwo view constraint;2) Simply by adding another template view using two viewconstraint cannot obtain much gain;3) A naive solution of using trifocal tensorinstead of the subtensor will lose useful information and lead to a deficiency of theconstraint. Experiments and datasets are designed to demonstrate the effectivenessand improved performance of the proposed approach. |