Beyond attributes: High-order attribute features for zero-shot learning

In this paper, SeeNet with the high-order attribute features (SeeNet-HAF) is proposed to solve the challenging zero-shot learning (ZSL) task. The high-order attribute features aims to discover a more elaborate, discriminative high-order semantic vector for each class and can distill the correlation between the class attributes embedding into modeling. SeeNet-HAF consists of two branches. The upper stream is capable of dynamically localizing some discriminative object region, and then the high-order attribute supervision is incorporated to characterize the relationship between the class attributes. Meanwhile, the bottom stream discovers complementary object regions by erasing its discovered regions from the feature maps. In addition, we propose a fast hyperparameter search strategy. It takes both the breadth and precision of the search into account. Experiments on four standard benchmark datasets demonstrate the superiority of the SeeNet-HAF framework.