Learning transferable distance functions for human action recognition and detection

In this thesis, we address an important topic in computer vision, human action recognition and detection. In particular, we focus on a special scenario where only a single clip is available for training for each action category. This is a very natural scenario in many real-world applications, such as video search and intelligent video surveillance. We present a transfer learning technique called transferable distance function learning and apply it in human action recognition and detection. This learning algorithm aims to extract generic knowledge from previous training sets, and apply this knowledge to videos of new actions without further learning. It is experimentally demonstrated that the proposed algorithm can improve the accuracy of single clip action recognition and detection. Based on the learned transferable distance function, we further propose a cascade structure which can significantly improve the efficiency of an action detection system.