Equiangular Aligned Dual Prompt Learning for Open-Set Recognition

Qifei Li, Enhao Zhang, Dong Liang, Chuanxing Geng

Abstract: Existing open-set recognition (OSR) methods based on pre-trained Vision-Language Models (VLMs) primarily focus on learning good semantic prompts to establish compact decision boundaries for known classes. Despite achieving promising results, such decision boundaries are insufficient to some extent due to the inherent gap between semantic and visual features. To this end, this paper proposes a novel Equiangular Aligned Dual Prompt Learning framework (EADPL) to alleviate such a gap. Specifically, alongside the existing semantic prompts, a visual prototype prompt module is innovatively proposed to compensate for the deficiencies of semantic prompts. Moreover, we also strategically introduce simplex equiangular alignment constraints for semantic prompts and visual prototype prompts separately, which not only maintains maximal inter-class separation for known classes to enhance their classification accuracy but also reserves more space for unknown classes. In particular, by constraining both semantic prompts and visual prototype prompts onto the same simplex geometric structure, the gap between semantic and visual features is further alleviated. Extensive experiments demonstrate the superiority of our EADPL.

External IDs:dblp:conf/icic/LiZLG25