Latent Fine-grained Features embedding model for Unsupervised Zero-shot Learning

Zero-shot learning (ZSL) aims to identify image categories that never appear in the training set by learning a mapping from visual information to the semantic space. For existing works in recent years, it has been paying much attention to user-defined-attributes-based supervised ZSL models and manual annotation-based unsupervised ZSL methods, whilst the important drawback of user-defined attributes is ignored. User-defined attributes usually need to be precisely annotated not only for seen classes but also for unseen classes. This procedure of collecting user-defined attributes is an error-prone and time-consuming work limiting the scalability of the methods to a great extent. Moreover, user-defined attributes are semantic embedding but they are not exhaustive. In this paper, we propose a manual annotation-free unsupervised ZSL method with a great scalability which is a benefit for the large-scale ZSL tasks, just using public word vectors of categories without any dedicated attributes annotation efforts. In addition, our approach can automatically extract latent fine-grained features to reduce visual information losses caused by the absence of user-defined attributes. Extensive experimental results show that our method outperforms the previous methods among the manual annotation-free unsupervised ZSL methods.