Category Discovery: An Open-World Perspective

2025-09-26

AI Summary

This paper addresses Open-World Class Discovery (OWCD), the problem of automatically clustering unlabeled data containing both seen and unseen classes, given only labeled data from known classes. We propose the first unified OWCD taxonomy, encompassing realistic settings such as novel-class discovery and generalized OWCD. Our analysis systematically identifies the critical roles of large-scale pretrained backbones, hierarchical representation learning, auxiliary information integration, and curriculum-style training. Core technical contributions include: (i) hierarchical and self-supervised representation learning; (ii) a robust label assignment mechanism mitigating pseudo-label noise; and (iii) a scalable approach to estimating the number of classes. We further explore distributed paradigms—including federated learning—to support complex, real-world deployments. Through comprehensive evaluation of state-of-the-art methods, we distill effective design principles and identify persistent challenges: label assignment bias, inaccurate class cardinality estimation, and multi-objective optimization trade-offs. This work establishes a theoretical framework and practical guidelines for advancing OWCD research.

Technology Category

• Machine Learning: Multi-class/Multi-label Learning & Extreme Classification
• Search and Optimization: Learning to Search
• Computer Vision: Representation Learning for Vision

Application Category

• Semantics and Knowledge: Scalable techniques for the creation, curation, publication, maintenance, and consumption of large, Web-based, structured, reusable, knowledge graphs and ontologies
• Web Mining and Content Analysis: Normalization, clustering, classification, and summarization of Web text
• Graph Algorithms and Modeling for the Web: Representation, reconstruction, and subgraph or motif discovery in Web-related graphs

Abstract

Category discovery (CD) is an emerging open-world learning task, which aims at automatically categorizing unlabelled data containing instances from unseen classes, given some labelled data from seen classes. This task has attracted significant attention over the years and leads to a rich body of literature trying to address the problem from different perspectives. In this survey, we provide a comprehensive review of the literature, and offer detailed analysis and in-depth discussion on different methods. Firstly, we introduce a taxonomy for the literature by considering two base settings, namely novel category discovery (NCD) and generalized category discovery (GCD), and several derived settings that are designed to address the extra challenges in different real-world application scenarios, including continual category discovery, skewed data distribution, federated category discovery, etc. Secondly, for each setting, we offer a detailed analysis of the methods encompassing three fundamental components, representation learning, label assignment, and estimation of class number. Thirdly, we benchmark all the methods and distill key insights showing that large-scale pretrained backbones, hierarchical and auxiliary cues, and curriculum-style training are all beneficial for category discovery, while challenges remain in the design of label assignment, the estimation of class numbers, and scaling to complex multi-object scenarios.Finally, we discuss the key insights from the literature so far and point out promising future research directions. We compile a living survey of the category discovery literature at href{https://github.com/Visual-AI/Category-Discovery}{https://github.com/Visual-AI/Category-Discovery}.

Problem

• Automatically categorizing unlabeled data from unseen classes
• Using labeled data from seen classes to discover new categories
• Addressing challenges in representation learning and class estimation

Innovation

• Large-scale pretrained backbones for representation learning
• Hierarchical and auxiliary cues for enhanced discovery
• Curriculum-style training for improved category discovery