Partwise bag-of-words-based multi-task learning for human action recognition

Proposed is a human action recognition method by partwise bag-of-words (BoW)-based multi-task learning. The authors present partwise BoW representation and furthermore formulate the action recognition task as a joint multi-task learning problem by transfer learning penalised by a graph structure and sparsity to discover latent correlation and boost performances. A large-scale experiment shows that this method can significantly improve performance over the standard BoW + SVM method. Moreover, the proposed method can achieve competing performances against the state-of-the-art methods for human action recognition in an effective and easy to follow way.