Search papers, labs, and topics across Lattice.
This paper introduces Compositional Primitive Fields (CPF-GCD), a novel framework designed to enhance Generalized Category Discovery (GCD) by reshaping the feature space to reveal latent structures. By enforcing a low-rank compositional organization, CPF-GCD transforms high-rank, entangled token representations from standard vision backbones into structured and separable primitive fields. Experimental results show that CPF-GCD significantly improves performance across various GCD baselines, highlighting the importance of low-rank compositional structures in enabling the discovery of novel categories in open-world settings.
Discovering new categories in open-world settings hinges on reshaping feature spaces, and CPF-GCD reveals how low-rank compositional structures can unlock this potential.
Generalized Category Discovery (GCD) aims to recognize known classes while autonomously discovering novel ones in open-world settings. However, current approaches primarily focus on designing clustering objectives, often overlooking a critical bottleneck: standard vision backbones yield high-rank, entangled token representations that are ill-suited for unsupervised discovery of latent concepts and structures. In this paper, we propose Compositional Primitive Fields (CPF-GCD), a novel representation learning framework that reshapes the feature space to make such latent structure identifiable by enforcing a low-rank compositional organization. Our core hypothesis is that all categories, whether known or novel, can be expressed as compositions and spatial arrangements of a finite set of learnable visual primitives that capture reusable concepts. CPF instantiates this geometric constraint via a spatial field mechanism. Inserted between the backbone and the head, it rewrites noisy patch tokens through low-rank primitive mixtures, effectively decomposing images into reusable atomic parts and their spatial layouts. By explicitly modeling the spatial distribution of primitives, CPF enables novel categories to emerge naturally as new activation patterns over a shared vocabulary. This shifts the focus of representation from merely partitioning global embeddings to constructing a structured and separable primitive field. Extensive experiments demonstrate that CPF serves as a generic, plug-and-play module that consistently boosts performance across diverse GCD baselines, validating that identifying and leveraging low-rank compositional structure is a crucial inductive bias for open-world recognition.
