Implicit and Explicit Attention for Zero-Shot Learning

Most of the existing Zero-Shot Learning (ZSL) methods focus on learning a compatibility function between image representation and class attributes. Few others concentrate combining local global features. However, approaches still fail to address bias issue towards seen classes. In this paper, we propose implicit explicit attention mechanisms problem in ZSL models. We formulate mechanism with self-supervised angle rotation task, which focuses specific features aiding solve task. The is composed consideration multi-headed self-attention via Vision Transformer model, learns map semantic space during training stage. conduct comprehensive experiments three popular benchmarks: AWA2, CUB SUN. performance our proposed has proved its effectiveness, achieved state-of-the-art harmonic mean all datasets.