Category Discovery:
An Open-World Perspective

Zhenqi He, Yuanpei Liu, Kai Han
Visual AI Lab, The University of Hong Kong

Paper arXiv Paper List BibTeX
Comparison between semi-supervised learning, open-set recognition and category discovery

Category Discovery groups unlabelled data into meaningful categories, guided by labelled known classes. Unlike semi-supervised learning, it assumes no fixed label space; unlike open-set or OOD detection, it discovers the structure of unknowns rather than just rejecting them. This work summarises and analyses the field's settings, methods, benchmarks, and open challenges from an open-world perspective.

🌱 A Living Repository

Help Keep This Repository Growing

Category discovery moves fast. We maintain an open, continuously updated repository of papers, benchmarks, and code — and it grows with the community. Spotted a new paper or a missing reference? Open a pull request and add it.

Contribute a Paper Browse the repository →

Core Contributions

A map of the field — from problem definitions to open-world scenarios and standardized benchmarks.

1

A Multi-Faceted Taxonomy

Two base settings, seven derived settings, and extensions to other modalities — showing how assumptions reshape the problem.

2

A Unified Method Lens

Every method reduces to three components: representation learning, label assignment, and class-number estimation.

3

Consolidated Benchmarks

NCD and GCD compared on generic and fine-grained datasets, with consistent backbones, splits, and protocols.

4

Actionable Research Directions

What works today, and the key gaps to close before reliable open-world discovery.

A Multi-Faceted Taxonomy

No single axis captures the field. We separate label-space assumptions, deployment challenges, and data modalities.

Two Base Settings

Novel Category Discovery (NCD)

Cluster an unlabelled set of only unseen categories, transferring knowledge from labelled base classes.

YL ∩ YU = ∅

Generalized Category Discovery (GCD)

Cluster unlabelled data mixing seen and unseen categories — more realistic, but biased toward labelled classes.

YL ∩ YU ≠ ∅

Derived Settings

Continual CD On-the-Fly CD CD with Domain Shift Distribution-Agnostic CD Semantic CD Few-Shot CD Federated CD

Beyond Visual Recognition

Segmentation & Detection Video 3D Point Clouds Text Tabular Data Graphs Multi-Modal Data
Overview of the category discovery taxonomy for base and derived settings

The taxonomy across two base settings and seven derived settings.

BibTeX

@article{he2026category,
  title={Category Discovery: An Open-World Perspective},
  author={He, Zhenqi and Liu, Yuanpei and Han, Kai},
  journal={arXiv preprint arXiv:2509.22542},
  year={2026},
  url={https://arxiv.org/abs/2509.22542}
}