In this paper, we present CymNet, a new approach for compositional zero-shot learning (CZSL) that builds upon the successful foundation of SymNet and introduces the use of CLIP to enhance its compatibility with zero-shot settings. Our research was motivated by the recognition that the potential of SymNet, which is based on the principles of symmetry and group theory, has not been fully exploited in CZSL tasks. By incorporating CLIP into SymNet, CymNet demonstrates improved performance on the UT-Zappos50K dataset, with a top-3 accuracy increase of 7%. We also conducted additional experiments to evaluate the necessity of various components of the model, including the CLIP adapter.


       Jiaming Shan        Zhaozi Wang       Mingshu Zhai


Our paper is available here.

How CymNet Works