Robust relative attributes for human action recognition

High-level semantic feature is important to recognize human action. Recently, relative attributes, which are used to describe relative relationship, have been proposed as one of high-level semantic features and have shown promising performance. However, the training process is very sensitive to noises and moreover it is not robust to zero-shot learning. In this paper, to overcome these drawbacks, we propose a robust learning framework using relative attributes for human action recognition. We simultaneously add Sigmoid and Gaussian envelops into the loss objective. In this way, the influence of outliers will be greatly reduced in the process of optimization, thus improving the accuracy. In addition, we adopt Gaussian Mixture models for better fitting the distribution of actions in rank score space. Correspondingly, a novel transfer strategy is proposed to evaluate the parameters of Gaussian Mixture models for unseen classes. Our method is verified on three challenging datasets (KTH, UIUC and HOLLYWOOD2), and the experimental results demonstrate that our method achieves better results than previous methods in both zero-shot classification and traditional recognition task for human action recognition.