Multi-Layer Cross Loss Model for Zero-Shot Human Activity Recognition

Wu, Tong; Chen, Yiqiang; Gu, Yang; Wang, Jiwei; Zhang, Siyu; Zhechen, Zhanghu

Most existing methods of human activity recognition are based on supervised learning. These methods can only recognize classes which appear in the training dataset, but are out of work when the classes are not in the training dataset. Zero-shot learning aims at solving this problem. In this paper, we propose a novel model termed Multi-Layer Cross Loss Model (MLCLM). Our model has two novel ideas: (1) In the model, we design a multi-nonlinear layers model to project features to semantic space for that the deeper the network is, the better the network can fit the data’s distribution. (2) A novel objective function combining mean square loss and cross entropy loss is designed for the zero-shot learning task. We have conduct sufficient experiments to evaluate the proposed model on three benchmark datasets. Experiments show that our model outperforms other state-of-the-art methods significantly in zero-shot human activity recognition.