A Hierarchical-Tree-Based Method for Generative Zero-Shot Learning

It is currently a popular practice to use the class semantic information and the conditional generative adversarial network (CGAN) technique to generate visual features for the unseen classes in zero-shot learning (ZSL). However, there is currently no good ways to ensure that the generated visual features can always be beneficial to the prediction of the unseen classes. To alleviate this problem, we propose a hierarchical-tree-based method for constraining the generation process of CGAN, which can tune the generated visual features based on the multi-level class information. Moreover, to enhance the mapping ability of the model from the visual space to the semantic space, we add a multi-expert module to the traditional single mapping channel, which helps the model to mine the mapping relationship between the visual space and the semantic space. Extensive experimental results on five benchmark data sets show that our method can achieve better generalization ability than other existing generative ZSL algorithms.