Multidomain Features Fusion for Zero-Shot Learning

Given a novel class instance, the purpose of zero-shot learning (ZSL) is to learn a model to classify the instance by seen samples and semantic information transcending class boundaries. The difficulty lies in how to find a suitable space for zero-shot recognition. The previous approaches use semantic space or visual space as classification space. These methods, which typically learn visual-semantic or semantic-visual mapping and directly exploit the output of the mapping function to measure similarity to classify new categories, do not adequately consider the complementarity and distribution gap of multiple domain information. In this paper, we propose to learn a multidomain information fusion space by a joint learning framework. Specifically, we consider the fusion space as a shared space in which different domain features can be recovered by simple linear transformation. By learning a n-way classifier of fusion space from the seen class samples, we also obtain the discriminative information of the similarity space to make the fusion representation more separable. Extensive experiments on popular benchmark datasets manifest that our approach achieves state-of-the-art performances in both supervised and unsupervised ZSL tasks.