Reproduce “Context-based and Diversity-driven Specificity in Compositional Zero-Shot Learning” on OpenResearch

About this paper

Yun Li, Zhe Liu, Hang Chen, Lina Yao

A new framework, CDS-CZSL, quantifies attribute informativeness through object cluster diversity to address attribute specificity and open-world scalability in compositional zero-shot learning. It establishes new state-of-the-art performance across benchmark datasets, achieving up to 3.2% AUC improvement on UT-Zappos in closed-world settings and robust generalization in open-world scenarios.

Key insights

• Attribute specificity, defined by the descriptive diversity across object clusters, is a critical and previously under-explored dimension that enhances compositional zero-shot learning accuracy.
• A unified framework combining composition-wise (contextual) and primitive-wise (disentangled) learning, enhanced with specific prompt designs and visual adapters, achieves superior performance in CZSL.
• Intelligent filtering based on learned attribute specificity is more effective for Open-World CZSL than similarity-based methods, preventing both search space explosion and loss of genuinely specific unseen compositions.

Summary by alphaXiv.