GNN-based primitive recombination for compositional zero-shot learning

Abstract

Compositional Zero-Shot Learning (CZSL) aims to recognize unseen attribute–object combinations, with the core challenge being the complex visual manifestations across compositions. We posit that the key to address this challenge lies in enabling models to simulate human recognition processes by decomposing and dynamically recombining primitives (attributes and objects). Existing methods merely concatenate primitives after extraction to form new combinations, without achieving deep integration between attributes and objects to create truly novel compositions. To address this issue, we propose Graph Neural Network-based Primitive Recombination (GPR) framework. This framework innovatively designs a Primitive Recombination Module (PRM) based on the Compositional Matching Module (CMM). Specifically, we first extract primitives, and build independent attribute and object space based on the CLIP model, enabling more precise learning of primitive-level visual features and reducing information residuals. Additionally, we introduce a Virtual Composition Unit (VCU), which inputs optimized primitive features as nodes into GNN and models complex interaction relationships between attributes and objects through message propagation. The module performs mean pooling on the updated node features to obtain a recombined representation and fuses the global visual information from the original image through residual connections, generating semantically rich virtual compositional features while preserving key visual cues. We conduct extensive experiments on three CZSL benchmark datasets to show that GPR achieves state-of-the-art or competitive performance in both closed-world and open-world settings.