Isolation and distillation network for generalized zero-shot learning

Generalized zero-shot learning (GZSL) aims to identify the samples that belong to a different label space than training samples, while the samples of these unseen classes have the same semantic space as the samples of seen classes. Therefore, most GZSL models use the samples to train the fundamental part, then use semantic information to match or generate the features of the unseen classes as the generalized part. However, the importance of features and their consistency with semantics are different. This difference causes jeopardizing between each other, becoming one of the main reasons that the classification results of seen and unseen classes in the GZSL are lower than the traditional classification and the CZSL. To maintain the training effect from two parts without jeopardizing each other, IDN-GZSL is presented in this paper. We provide rearrangement modules in different isolated scenarios to distinguish the positive and negative parts of the feature and extract the positive features. Three rearrangement modules have been trained using existing training data and generated data in corresponding scenarios. Although the three scenarios are isolated, the decoupling and relation modules are used to support the rearrangement and acquire positive features for semantic information, and the distillation module is used to establish indirect connections. The results demonstrate that our methods take effect with three significant improvements in ablation experiments: successfully trained rearrangement modules for obtaining positive features in multiple scenarios for classification, eliminated jeopardizing among different scenarios by establishing three isolation scenarios, and established necessary connections among them by distillation module. Our method performs the best classification ability of unseen classes on four datasets and achieves the best performance on AWA2, CUB, and FLO when evaluating the method through harmonic means.