Salient Latent Features For Zero-shot Learning

Zero-shot recognition (ZSR) refers to predict the categories of the unseen new images that never appear in the training set. Seen classes belong to the training set and unseen classes belong to the test set. Typically, most of the previous works focus on solving how to embed the space between visual and semantic representation. In this paper, we argue that the salience of the image can provide more discriminative information for ZSR. We propose a framework that first detects the salient objects in image, and then locates the most interested salient region and extracts the features in it. We use the convolutional neural network (CNN) to extract the visual features of the salient region and maps them to both user-defined and latent attributes. Ablation study on two challenging ZSL datasets is conducted. The experimental results demonstrate the salient latent features extracting from the salient origins can improve the recognition accuracy in a great deal.