Multi-Head Self-Attention via Vision Transformer for Zero-Shot Learning
It addresses the problem of recognizing unseen object classes in computer vision, with incremental improvements in benchmark performance.
The paper tackled zero-shot learning by proposing an attention-based model using Vision Transformer to learn discriminative attributes from image patches, achieving new state-of-the-art harmonic mean results on AWA2, CUB, and SUN benchmarks.
Zero-Shot Learning (ZSL) aims to recognise unseen object classes, which are not observed during the training phase. The existing body of works on ZSL mostly relies on pretrained visual features and lacks the explicit attribute localisation mechanism on images. In this work, we propose an attention-based model in the problem settings of ZSL to learn attributes useful for unseen class recognition. Our method uses an attention mechanism adapted from Vision Transformer to capture and learn discriminative attributes by splitting images into small patches. We conduct experiments on three popular ZSL benchmarks (i.e., AWA2, CUB and SUN) and set new state-of-the-art harmonic mean results {on all the three datasets}, which illustrate the effectiveness of our proposed method.