Discovering latent attributes for human action recognition in depth sequence

A method for human action recognition in depth sequence by discovering latent attributes is proposed. Specifically, a latent attribute model to take advantage of the partwise and bodywise action attributes and their concurrence for action modelling is presented. This model can avoid the explicit definition of a completed action attribute set by designing a hidden layer as latent attributes. Moreover, the relationship between pairwise attributes can be adaptively learned in terms of the specific graphical structure without human intervention. Extensive experiments on two popular datasets and a new dataset TJU show the superiority of the proposed method over the state-of-the-arts.