Generating Manifold-Aligned Semantic Feature for Zero-Shot Learning

Zero-shot learning (ZSL) is a task that requires us to recognize novel classes whose visual instances are not included in training set. Most of the state-of-the-art ZSL methods resort to semantic embeddings to learn a visual-semantic mapping. However, there exists a huge gap between image feature space and the semantic embedding space. The main difficulty lies in how to learn a mapping with strong transfer ability to unseen classes. Motivated by the strong domain transfer ability of Generative Adversarial Networks (GAN), we introduce a new method to generate new semantic embeddings, whose manifold is designed to align with the manifold in the image feature space. These new semantic embeddings can boost the transfer ability of the visual-semantic mapping to unseen classes, thus leading to the improvement of ZSL performance. Experimental results on benchmarks show the superiority over the state-of-the-arts.