Multi-label zero-shot human action recognition via joint latent ranking embedding.

Abstract

Human action recognition is one of the most challenging tasks in computer vision. Most of the existing works in human action recognition are limited to single-label classification. A real-world video stream, however, often contains multiple human actions. Such a video stream is usually annotated collectively with a set of relevant human action labels, which leads to a multi-label learning problem. Furthermore, there are a great number of meaningful human actions in reality but it would be extremely difficult, if not impossible, to collect/annotate sufficient video clips regarding all these human actions for training a supervised learning model. In this paper, we formulate a real-world human action recognition task as a multi-label zero-shot learning problem. To address this problem, a joint latent ranking embedding framework is proposed. Our framework holistically tackles the issue of unknown temporal boundaries between different actions within a video clip for multi-label learning and exploits the side information regarding the semantic relationship between different human actions for zero-shot learning. Specifically, our framework consists of two component neural networks for visual and semantic embedding respectively. Thus, multi-label zero-shot recognition is done by measuring relatedness scores of concerned action labels to a test video clip in the joint latent visual and semantic embedding spaces. We evaluate our framework in different settings, including a novel data split scheme designed especially for evaluating multi-label zero-shot learning. The experimental results on two weakly annotated multi-label human action datasets (i.e. Breakfast and Charades) demonstrate the effectiveness of our framework.