Simple Is Better: A Global Semantic Consistency Based End-to-End Framework for Effective Zero-Shot Learning

In image recognition, there are many cases where training samples cannot cover all target classes. Zero-shot learning (ZSL) addresses such cases by classifying the samples of unseen categories that have no corresponding samples contained in the training set via class semantic information. In this paper, we propose a novel and simple end-to-end framework, called Global Semantic Consistency Network (GSC-Net for short), which makes complete use of the semantic information of both seen and unseen classes to support effective zero-shot learning. We also employ a soft label embedding loss to further exploit the semantic relationships among classes and use a seen-class weight regularization to balance attribute learning. Moreover, to adapt GSC-Net to the setting of Generalized Zero-shot Learning (GZSL), we introduce a parametric novelty detection mechanism. Experiments on all the three widely-used ZSL datasets show that GSC-Net performs better than most existing methods under both ZSL and GZSL settings. Especially, GSC-Net achieves the state of the art performance on two datasets (AWA2 and CUB). We explain the effectiveness of GSC-Net from the perspectives of class attribute learning and visual feature learning, and discover that the validation accuracy of seen classes can serve as an indicator of ZSL performance.