Attribute relaxation from class level to instance level for zero-shot learning

Zhang, Haofeng; Long, Yang; Zhao, Chunxia

Abstract

Conventional zero-shot learning (ZSL) methods usually use class-level attribute, which corresponds to a batch of images of same category. This setting is not reasonable since the images even though belong to same category still have variances in their attribute items. To alleviate this phenomenon, the authors propose a novel method namely attribute relaxation (AR) to extend attributes from class level to instance level by adding a small variance matrix, which is more reasonable than traditional ZSL methods such as Semantic AutoEncoder that projects features from multi to one. Extensive experiments on four popular datasets show that AR can significantly improve the method using only class-level attributes, and verifies that AR can make the projected features in attribute space more discriminative.

Citation

Zhang, H., Long, Y., & Zhao, C. (2018). Attribute relaxation from class level to instance level for zero-shot learning. Electronics Letters, 54(20), 1170-1172. https://doi.org/10.1049/el.2018.5027

• Journal Article Type: Article
• Online Publication Date: Oct 1, 2018
• Publication Date: 2018-10
• Deposit Date: Aug 31, 2019
• Journal: Electronics Letters
• Print ISSN: 0013-5194
• Publisher: Institution of Engineering and Technology (IET)
• Peer Reviewed: Peer Reviewed
• Volume: 54
• Issue: 20
• Pages: 1170-1172
• DOI: https://doi.org/10.1049/el.2018.5027
• Public URL: https://durham-repository.worktribe.com/output/1294119