Action Recognition Based on Latent Task Learning

In this paper, a novel approach based on latent task learning was presented for action recognition. A set of sub-actions shared by different actions were taken as the latent task, and then the latent task were jointly learned to model the intrinsic relationship among multiple actions for classifier training and the action recognition of video person. Specifically, a softmax based multi-class model was introduced to learn the latent tasks to avoid the ambiguity during recognition process and save the training time owing to its simple computation. Experimental results on UCF sports and Olympic sports datasets show that, compared with the binary-class based multi-task method the proposed method not only saves the running time from 130 s per iteration to 0.5 s per iteration, but also achieves better performance on action recognition tasks.