Complementary Attributes: A New Clue to Zero-Shot Learning

Zero-shot learning (ZSL) aims to recognize unseen objects using disjoint seen via sharing attributes. The generalization performance of ZSL is governed by the attributes, which transfer semantic information from classes classes. To take full advantage knowledge transferred in this paper, we introduce notion complementary attributes (CAs), as a supplement original enhance representation ability. Theoretical analyses demonstrate that CAs can improve PAC-style bound model. Since proposed CA focuses on enhancing representation, be easily applied any existing attribute-based methods, including label-embedding strategy-based (LEZSL) and probability-prediction (PPZSL). In PPZSL, there strong assumption all are independent each other, arguably unrealistic practice. solve problem, novel rank aggregation (RA) framework circumvent assumption. Extensive experiments five benchmark datasets large-scale ImageNet dataset RA significantly robustly methods achieve state-of-the-art performance.