Kernelized distance learning for zero-shot recognition

Abstract

Zero-Shot Learning (ZSL) has gained growing attention over the past few years mostly because it provides a significant scalability to recognition models for classifying instances from new unobserved classes. This scalability is achieved by providing semantic information about new classes, which could be obtained remarkably easier with lower cost, compared to collecting a new training set. Because seen and unseen classes are completely disjoint, ZSL methods often suffer from domain shift problem that occurs in transferring the knowledge of seen classes to unseen ones. Moreover, hubness problem that usually arises in high-dimensional space is another challenge in most ZSL methods due to applying nearest neighbor search for classification. To address these issues, a kernelized distance function is learned in order to discriminate the classes with a customized large-margin loss function. Furthermore, a simple theoretical-based prototype learning approach is provided by defining a non-linear mapping function to learn the visual prototype of each class from associated semantic information. For classification task, the learned distance function is utilized to measure the distance between instances and class-related prototypes. The evaluation on five benchmarks demonstrates the superiority of the proposed method over the state-of-the-art approaches in both zero-shot and generalized zero-shot learning problems.