Mutual Balancing in State-Object Components for Compositional Zero-Shot Learning

• Published in: Pattern recognition 2024-08, Vol.152, p.110451, Article 110451
• Main Authors: Jiang, Chenyi, Ye, Qiaolin, Wang, Shidong, Shen, Yuming, Zhang, Zheng, Zhang, Haofeng
• Format: Article
• Language: English
• Subjects: Compositional Zero-Shot Learning; Image classification; Mutual Balancing; Visual-attribute
• ISSN: 0031-3203; 1873-5142
• DOI: 10.1016/j.patcog.2024.110451

Compositional Zero-Shot Learning (CZSL) aims to recognize unseen compositions from seen states and objects. The disparity between the manually labeled semantic information and its actual visual features causes a significant imbalance of visual deviation in the distribution of various object classes and state classes, which is ignored by existing methods. To ameliorate these issues, we consider the CZSL task as an unbalanced multi-label classification task and propose a novel method called MUtual balancing in STate-object components (MUST) for CZSL, which provides a balancing inductive bias for the model. In particular, we split the classification of the composition classes into two consecutive processes to analyze the entanglement of the two components to get additional knowledge in advance, which reflects the degree of visual deviation between the two components. We use the knowledge gained to modify the model’s training process in order to generate more distinct class borders for classes with significant visual deviations. Extensive experiments demonstrate that our approach significantly outperforms the state-of-the-art on MIT-States, UT-Zappos, and C-GQA when combined with the basic CZSL frameworks, and it can improve various CZSL frameworks. Our code is available at https://github.com/LanchJL/MUST. •The method considers CZSL as an unbalanced multi-label classification, utilizing visual deviation of components to provide an inductive bias.•Component imbalance info is used to re-weight CZSL training, enabling the model to reconstruct inter-component balance.•The method outperforms SoTAs with base CZSL methods, and augments joint embedding function based approaches.