Consistent Supervised-Unsupervised Alignment for Generalized Category Discovery

TL;DR

NC-GCD addresses inconsistent objectives and category confusion in Generalized Category Discovery by fixing a Simplex ETF prototype space and enforcing consistent alignment between supervised labels and unsupervised clusters. A Consistent ETF Alignment Loss combines supervised and unsupervised ETF alignment, while the Semantic Consistency Matcher stabilizes cluster-label mappings across iterations, yielding a Geometry-driven, NC-inspired feature space for both known and novel categories. Across six benchmarks and multiple backbones, NC-GCD achieves state-of-the-art performance, especially in novel-category accuracy, and demonstrates robustness to K estimation and clustering dynamics. The work provides a principled, efficient framework for open-world category learning with strong practical impact for scalable recognition and discovery.

Abstract

Generalized Category Discovery (GCD) focuses on classifying known categories while simultaneously discovering novel categories from unlabeled data. However, previous GCD methods face challenges due to inconsistent optimization objectives and category confusion. This leads to feature overlap and ultimately hinders performance on novel categories. To address these issues, we propose the Neural Collapse-inspired Generalized Category Discovery (NC-GCD) framework. By pre-assigning and fixing Equiangular Tight Frame (ETF) prototypes, our method ensures an optimal geometric structure and a consistent optimization objective for both known and novel categories. We introduce a Consistent ETF Alignment Loss that unifies supervised and unsupervised ETF alignment and enhances category separability. Additionally, a Semantic Consistency Matcher (SCM) is designed to maintain stable and consistent label assignments across clustering iterations. Our method achieves strong performance on multiple GCD benchmarks, significantly enhancing novel category accuracy and demonstrating its effectiveness.

Figures (12)

• Figure 1: Illustration of the differences between prior studies, our proposed NC-GCD framework, and Performance Overview. (a) Our method pre-assigns fixed ETF prototypes rather than dynamically learning prototypes, ensuring consistent optimization objectives for known and novel categories. (b) The overview of the NC-GCD. (c) Compared to previous studies, our method exhibits superior overall accuracy, with a notable improvement in the accuracy of novel categories.
• Figure 2: Overview of the NC-GCD Framework. The framework uses periodic clustering to group features, aligning them with pre-assigned ETF prototypes. The Semantic Consistency Matcher (SCM) ensures consistent label assignments across clustering iterations. This process stabilizes feature alignment and maintains a consistent geometric structure for both known and novel categories.
• Figure 3: Ablation study of the $\alpha$ on CUB-200.
• Figure 4: Ablation study of our SCM module.
• Figure 4: Visualization of Comparison with DINOv1, CMS, and our method on CUB-200.