ProCC: Progressive Cross-Primitive Compatibility for Open-World Compositional Zero-Shot Learning

• DOI: 10.1609/aaai.v38i11.29164
• Publication Date: 2024-03-25T10:53:58Z
• Authors: Fushuo Huo; Wenchao Xu; Song Guo; Jingcai Guo; Haozhao Wang; Ziming Liu; Xiaocheng Lu
• Keywords: FOS: Computer and information sciences; 0303 health sciences; 03 medical and health sciences; Computer Vision and Pattern Recognition (cs.CV); Computer Science - Computer Vision and Pattern Recognition

Open-World Compositional Zero-shot Learning (OW-CZSL) aims to recognize novel compositions of state and object primitives in images with no priors on the compositional space, which induces a tremendously large output space containing all possible state-object compositions. Existing works either learn joint embedding or predict simple separate classifiers. However, former method heavily relies external word methods, latter ignores interactions interdependent primitives, respectively. In this paper, we revisit primitive prediction approach propose method, termed Progressive Cross-primitive Compatibility (ProCC), mimic human learning process for OW-CZSL tasks. Specifically, cross-primitive compatibility module explicitly learns model features trainable memory units, efficiently acquires visual attention reason high-feasibility compositions, without aid knowledge. Moreover, alleviate invalid interactions, especially partial-supervision conditions (pCZSL), design progressive training paradigm optimize classifiers conditioned pre-trained an easy-to-hard manner. Extensive experiments three widely used benchmark datasets demonstrate that our outperforms other representative methods both pCZSL settings by margins.