Category Discovery: An Open-World Perspective

Zhenqi He , Yuanpei Liu , Kai Han

This repository serves as a supplementary resource for our survey paper on Category Discovery (CD) methods. It includes a comprehensive collection of key papers, frameworks, and approaches in the field of CD, summarizing the most recent advancements and techniques. The materials here aim to provide researchers with an accessible overview of current trends and methodologies in CD, along with references and additional insights to support further exploration.

We will continue to maintain and update this repository with new papers and resources as the field evolves. Contributions are welcome, and we encourage pull requests (PRs) to help expand and improve the content for the community.

Table of Contents

• Introduction
• Category Discovery
  • Novel Category Discovery
  • Generalized Category Discovery
  • Continual Category Discovery
  • On-the-fly Category Discovery
  • Category Discovery with domain shift
  • Distribution-Agnostic Category Discovery
  • Semantic Category Discovery
  • Few-Shot Category Discovery
  • Federated Category Discovery

Introduction

Category Discovery (CD) addresses the limitations of the closed-world assumption by embracing an open-world setting. As shown in above figure, CD differs from semi-supervised learning and OSR&OOD by clustering unlabelled data that contains unseen categories. It is motivated by that human beings are capable of discovering unknown species by transferring existing knowledge on explored species. CD proves highly applicable across various real-world scenarios. For example, in autonomous driving, vehicles must continuously detect and classify new objects—such as unfamiliar road signs or obstacles—beyond their initial training to ensure safe navigation. In retail, CD can automatically recognize newly introduced products in supermarkets without the need for manual labeling.

Roadmap

In recent years, CD has garnered increasing attention, leading to a proliferation of research exploring various methodologies and settings. It was initially introduced as Novel Category Discovery (NCD) in to cluster unlabelled novel categories by leveraging knowledge from labelled base categories. This concept was later expanded into Generalized Category Discovery (GCD), which relaxed earlier constraints by assuming that the unlabelled data contains both novel and base categories, thereby more closely mirroring real-world scenarios. Further advancing the field, Han *etal.* proposed Semantic Category Discovery (SCD), aiming to assign semantic labels to unlabelled samples from an unconstrained vocabulary space. Additionally, CD has been applied to complex scenarios such as continual learning, where models learn incrementally over time, and federated learning, which focuses on training models across decentralized devices while ensuring data privacy. CD methods have also been explored in challenging settings, including few-shot learning, where limited labelled data is available, and with imbalanced distribution and domain-shifted data, making CD more applicable to real-world problems.

Category Discovery

Novel Category Discovery (NCD)

The concept of NCD aims to transfer the knowledge learned from base categories to cluster unlabelled unseen categories, motivated by the observation where a child could easily distinguish novel categories (e.g., birds and elephants) after learning to classify base categories (e.g., dogs and cats).

Formally, given a dataset $`\mathcal{D} = \mathcal{D}_L \cup \mathcal{D}_U`$, where the labelled portion is $`\mathcal{D}_L = \{(\mathbf{x}_i, y_i)\}_{i=1}^M \subset \mathcal{X} \times \mathcal{Y}_L`$ and the unlabelled portion is $`\mathcal{D}_U = \{(\mathbf{x}_i, \hat{y}_i)\}_{i=1}^K \subset \mathcal{X} \times \mathcal{Y}_U`$ (with the labels $`\hat{y}_i`$ being inaccessible during training), the objective of NCD is to leverage the discriminative information learned from the annotated data to cluster the unlabelled data.

This setting presumes that the label spaces of the labelled and unlabelled data are disjoint, i.e., $\mathcal{Y}_L \cap \mathcal{Y}_U = \varnothing$, implying $\mathcal{C}_N = \mathcal{Y}_U$, while also assuming a high degree of semantic similarity between the base and novel categories.

Year	Method	Pub.	Backbone	Label Assignment	# Unlabelled categories	Dataset
2018	KCL	ICLR	ResNet	Parametric Classifier	Over-estimate	Omniglot, ImageNeg-1K, Office31
2019	MCL	ICLR	ResNet, VGG, LeNet	Parametric Classifier	Over-estimate	Omniglot, CIFAR-10&100, ImageNet-1K, MNIST
	DTC	ICCV	ResNet, VGG	Soft Assignment	$k$-Means	Omniglot, CIFAR-10&100, ImageNet-1K, SVHN
2020	RS, RS+	ICLR	ResNet	Parametric Classifier	Known	Omniglot, CIFAR-10&100, ImageNet-1K, SVHN
2021	Qing etal.	Neural Networks	ResNet	Parametric Classifier	Known	CIFAR-10&100, SVHN
	OpenMix	CVPR	ResNet, VGG	Parametric Classifier	Known	CIFAR-10&100, ImageNet-1K
	NCL	CVPR	ResNet	Parametric Classifier	Known	CIFAR-10&100, ImageNet-1K
	JOINT	ICCV	ResNet	Parametric Classifier	Known	CIFAR-10&100, ImageNet-1K
	UNO	ICCV	ResNet	Parametric Classifier	Known	CIFAR-10&100, ImageNet-1K
	DualRS	ICCV	ResNet	Parametric Classifier	Known	CIFAR-10&100, ImageNet-100&1K, SSB
2022	SMI	ICASSP	VGG-16	$k$-Means	Known	CIFAR-10&100, ImageNet-1K
	PSSCNNCD	T'CYB	N/A	BKBH $k$-Means	Progressive label propagation	Coil20, Yeast, MSRA25, PalmData25, Abalone, USPS, Letter, MNIST
	Li etal.	NeurIPSW	ResNet	$k$-Means	$k$-Means	CIFAR-100, ImageNet-1K
2023	ResTune	T'NNLS	ResNet	$k$-Means	Known	CIFAR-10&100, TinyImageNet
	SK-Hurt	TMLR	ResNet	$k$-Means	$k$-Means	CIFAR-100, ImageNet-1K
	IIC	CVPR	ResNet	Parametric Classifier	$k$-Means	CIFAR-10&100, ImageNet-1K
	NSCL	ICML	ResNet	$k$-Means	$k$-Means	CIFAR-100, ImageNet-1K
	CRKD	ICCV	ResNet, ViT	Parametric Classifier	Known	CIFAR-100, SSB
	Feng etal.	MICCAI	ResNet	Parametric Classifier	Known	ISIC2019
2024	RAPL	CVPR	ResNet	$k$-Means	Known	SoyAgeing
	SCKD	ECCV	ResNet, ViT	Parametric Classifier	Known	CIFAR-10&100, ImageNet-100, SSB
	APL	T'PAMI	ResNet	Parametric Classifier	Known	CIFAR-10&100, Omniglot, ImageNet-1K
PrePrint	Hasan etal.	ArXiv	ResNet	Parametric Classifier	$k$-Means	CIFAR-10&100

Generalized Category Discovery (GCD)

Extending the NCD paradigm, Generalized Category Discovery relaxes the disjointness assumption between the base and novel categories, thereby presenting a more challenging and realistic scenario. In GCD, the labelled and unlabelled datasets may share common categories, i.e., $\mathcal{Y}_L \cap \mathcal{Y}_U \neq \varnothing$, and the set of novel categories is defined as a subset of $\mathcal{Y}_U$ (i.e., $\mathcal{C}_N \subset \mathcal{Y}_U$). This general formulation is particularly pertinent to practical applications such as plant species discovery, where an existing database of known species is augmented with newly observed species, necessitating the clustering of both known and novel instances.

Notably, an equivalent formulation has been introduced by Cao etal. under the designation of Open-World Semi-Supervised Learning. In the following context, we refer to both formulations under the umbrella term Generalized Category Discovery.

Year	Method	Pub.	Backbone	Label Assignment	# Unlabelled categories	Dataset
2022	GCD	CVPR	ViT	Semi-k-Means	$k$-Means	CIFAR-10&100, ImageNet-100, SSB, Herb19
	ORCA	CVPR	ResNet	Parametric Classifier	Known	CIFAR-10&100, ImageNet-100, Single-Cell
	ComEx	CVPR	ResNet	Parametric Classifier	Known	CIFAR-10&100
	OpenLDN	ECCV	ResNet	Parametric Classifier	Known	CIFAR-10&100, ImageNet-100, TinyImage, Oxford Pets
	TRSSL	ECCV	ResNet	Parametric Classifier	$k$-Means	CIFAR-10&100, ImageNet-100, TinyImage, Oxford Pets, Scars, Aircrafts
	NACH	NeurIPS	ResNet	Parametric Classifier	Known	CIFAR-10&100, ImageNet-100
	XCon	BMVC	ViT	Semi-k-Means	$k$-Means	CIFAR-10&100, ImageNet-100, SSB, Oxford Pets
2023	OpenCon	TMLR	ResNet	Prototype-based	$k$-Means	CIFAR-10&100, ImageNet-100
	PromptCAL	CVPR	ViT	Semi-k-Means	Known	CIFAR-10&100, ImageNet-100, SSB
	DCCL	CVPR	ViT	Infomap	Infomap	CIFAR-10&100, ImageNet-100, CUB, Scars, Oxford Pets
	OpenNCD	IJCAI	ResNet	Prototype-based	Prototype Grouping	CIFAR-10&100, ImageNet-100
	SimGCD	ICCV	ViT	Parametric Classifier	$k$-Means	CIFAR-10&100, ImageNet-100, SSB, Herb19
	GPC	ICCV	ViT	GMM	GMM	CIFAR-10&100, ImageNet-100, SSB
	PIM	ICCV	ViT	Parametric Classifier	$k$-Means	CIFAR-10&100, ImageNet-100, CUB, Scars, Herb19
	TIDA	NeurIPS	ResNet	Parametric Classifier	Known	CIFAR-10&100, ImageNet-100, TinyImageNet, Scars, Aircraft
	$\mu$GCD	NeurIPS	ResNet, ViT, ViT	$k$-Means	Known	Clevr-4
	InfoSieve	NeurIPS	ViT	$k$-Means	$k$-Means	CIFAR-10&100, ImageNet-100, SSB, Oxford Pets, Herb19
	SORL	NeurIPS	ResNet	$k$-Means	Known	CIFAR-10&100
	Yang etal.	ICONIP	ViT	Louvain	Louvain	CIFAR-10&100, ImageNet-100, CUB, Scars, Herb19
2024	AMEND	WACV	ViT	Parametric Classifier	Known	CIFAR-10&100, ImageNet-100, SSB, Herb19
	GCA	WACV	ViT	Guided Cluster Aggregation	$k$-Means	CIFAR-10&100, ImageNet-100, SSB
	SPT-Net	ICLR	ViT, ViT	Parametric Classifier	$k$-Means	CIFAR-10&100, ImageNet-100, SSB
	LegoGCD	CVPR	ViT	Parametric Classifier	Known	CIFAR-10&100, ImageNet-100, -1K, SSB, Herb19
	CMS	CVPR	ViT	Agglomerative Clustering	Agglomerative Clustering	CIFAR-100, ImageNet-100, SSB, Herb19
	ActiveGCD	CVPR	ViT	Parametric Classifier	Known	CIFAR-10&100, ImageNet-100, SSB
	TextGCD	ECCV	ViT	Parametric Classifier	Known	CIFAR-10&100, ImageNet-100, -1K, SSB, Oxford Pets, Flowers102
	LPS	IJCAI	ResNet	Parametric Classifier	Known	CIFAR-10&100, ImageNet-100
	Contextuality-GCD	ICIP	ViT	Parametric Classifier	Known	CIFAR-10&100, ImageNet-100, -1K, SSB, Herb19
2025	MSGCD	Information Fusion	ViT	Parametric Classifier	Known	CIFAR-100, SSB
	CPT	IJCV	ViT	Similarity-Based	$k$-Means	CIFAR-10&100, ImageNet-100,CUB, Scars, Herb19
	PAL-GCD	AAAI	ViT	Parametric Classifier	DBSCAN	CIFAR-100,ImageNet-100,SSB, Herb19
	DebGCD	ICLR	ViT	Parametric Classifier	DBSCAN	CIFAR-10&100,ImageNet-100&1K,SSB, Herb19,Oxford-Pets
	ProtoGCD	T'PAMI	ViT	Parametric Classifier	$k$-Means	CIFAR-10&100,ImageNet-100&1K,SSB, Herb19
	MOS	CVPR	ViT	Parametric Classifier	Known	SSB, Oxford-Pets
	GET	CVPR	ViT	Parametric Classifier	Known	CIFAR-10&100,ImageNet-100,SSB,Herb19
	AptGCD	CVPR	ViT	Parametric Classifier	Known	CIFAR-10&100,ImageNet-100,SSB,Herb19
	Dai et al	CVPR	ViT	-	Known	SSB, Herb19
	HypCD	CVPR	ViT	-	Known	CIFAR-10&100,ImageNet-100,SSB,Herb19
PrePrint	CLIP-GCD	ArXiv	ViT	Semi-k-Means	$k$-Means	CIFAR-10&100, ImageNet-100, -1K, SSB, Flowers102, DomainNet
	MCDL	ArXiv	ViT	Parametric Classifier	Known	CIFAR-10&100, ImageNet-100, -1K, CUB, SCars, Herb19
	PNP	ArXiv	ViT	Infomap	Infomap	CIFAR-10&100, ImageNet-100, -1K, SSB, Herb19
	RPIM	ArXiv	ViT	Parametric Classifier	Known	CIFAR-10&100, ImageNet-100, -1K, CUB, Scars, Herb19
	OpenGCD	ArXiv	ViT	Parametric Classifier	$k$-Means	CIFAR-10&100, CUB
	ConceptGCD	ArXiv	ViT, ViT	Parametric Classifier	$k$-Means	CIFAR-100, ImageNet-100, -1K, SSB, Herb19
	GET	ArXiv	ViT	Parametric Classifier	Known	CIFAR-10&100, ImageNet-100, SSB, Herb19

Continual Category Discovery (CCD)

CCD provides a continual setting of category discovery in which new categories are identified sequentially while retaining previously acquired knowledge. CCD presents several distinct scenarios based on the structure of the incoming data.

In the Class Incremental Scenario, the training set $`\mathcal{D}_{\mathrm{train}}^t`$ contains solely unlabelled instances from novel categories. In the Mixed Incremental Scenario, $`\mathcal{D}_{\mathrm{train}}^t`$ is composed exclusively of unlabelled data drawn from both novel and base categories. Finally, in the Semi-Supervised Mixed Incremental Scenario, $`\mathcal{D}_{\mathrm{train}}^t`$ comprises both labelled and unlabelled samples, which originate from the base as well as the novel categories.

Year	Method	Pub.	Backbone	Scenario	Label Assignment	# Unlabelled categories	Dataset
2022	NCDwF	ECCV	ResNet	Class Incremental	Parametric Classifier	Known	CIFAR-10/100, ImageNet-1K
	FRoST	ECCV	ResNet	Class Incremental	Parametric Classifier	Known	CIFAR-10/100, TinyImageNet
	GM	NeurIPS	ResNet	All	Parametric Classifier	Known	CIFAR-100, ImageNet-100, CUB
2023	PA-GCD	ICCV	ViT, Resnet	Mixed Incremental	Parametric Classifier	Affinity Propagation	CUB, MIT67, Stanford Dogs, Aircraft
	MetaGCD	ICCV	ViT	Mixed Incremental	$k$-Means	$k$-Means	CIFAR-10/100, TinyImageNet
	iGCD	ICCV	ResNet	Self-Supervised Mixed Incremental	Soft Nearest Neighbor	Density Peaks	CUB, Aircraft, CIFAR-100
2024	Msc-iNCD	ICPR	ViT	Class Incremental	Parametric Classifier	Known	CIFAR-100, ImageNet-100/1K
	ADM	AAAI	ResNet	Class Incremental	Parametric Classifier	Known	CIFAR-10/100, TinyImageNet
	PromptCCD	ECCV	ViT	Mixed Incremental	GMM	GMP	CIFAR-100, ImageNet-100, TinyImageNet
	DEAN	ECCV	ViT	Mixed Incremental	Parametric Classifier	Affinity Propagation	CUB, Aircraft, CIFAR-100
	CAMP	ECCV	ViT	Self-Supervised Mixed Incremental	Nearest Centroid Classifier	Known	CUB, Aircraft, SCars, DomainNet, CIFAR-100
	Happy	NeurIPS	ViT	Mixed Incremental	Parametric Classifier	Silhouette Score	CIFAR-100, ImageNet-100, TinyImageNet, CUB
Preprint	FEA	ArXiv	ViT	Class Incremental	Parametric Classifier	Known	CIFAR-10/100, TinyImageNet

On-the-fly Category Discovery (OCD)

OCD extends conventional category discovery to an inductive learning paradigm with streaming inference. It trains on a labelled support set $D_S$ to cluster unlabelled query set $D_Q$ where $D_S$ is unavailable during training and its samples are individually at test time.

Year	Method	Pub.	Backbone	Label Assignment	# Unlabelled categories	Dataset
2023	SMILE	CVPR	ViT	Hash-based	Hash-coding	CIFAR-10&100, ImageNet-100, CUB, Scars, Herb19
2024	PHE	NeurIPS	ViT	Hamming Ball-Based	Hamming Ball-Based	CUB, Scars, Oxford Pets, Food-101, iNaturalist

Category Discovery with domain shift

This setting relaxes the conventional assumption that both labelled and unlabelled data are drawn from the same semantic domain. Formally, let $`\mathcal{D}_L`$ denote the labelled data, assumed to be exclusively drawn from the domain $`\Omega_B`$, and let $`\mathcal{D}_U`$ denote the unlabelled data, which may include samples originating from both $`\Omega_B`$ and an additional domain $`\Omega_{N}`$. The objective is to accurately classify images drawn from the combined domain $`\Omega = \Omega_B \cup \Omega_{N}`$, under the assumption that the novel domain is disjoint from the base domain (i.e., $`\Omega_B \cap \Omega_{N} = \varnothing`$). In practice, the novel domain $`\Omega_{N}`$ may encompass multiple subdomains.

Year	Method	Pub.	Backbone	$\Omega_{\mathcal{U}}$	Label Assignment	# Unlabelled categories	Dataset	$\mathcal{Y_L} \cap \mathcal{Y_U}$
2022	Yu etal.	AAAI	ResNet	Single New Domain	Parametric Classifier	$k$-Means	Office, OfficeHome, VisDA	$\varnothing$
	SCDA	ICME	ResNet	Multiple New Domains	Parametric Classifier	$k$-Means	Office, OfficeHome, DomainNet	$\varnothing$
2023	SAN	ICCV	ResNet	Single New Domain	Parametric Classifier	N/A	Office, OfficeHome, VisDA, DomainNet	$\varnothing$
2024	CDAD-Net	CVPRW	ViT	Single New Domain	Semi-k-Means	Elbow	OfficeHome, PACS, DomainNet, CIFAR-10&100, ImageNet-100	$\neq \varnothing$
2025	HiLo	ICLR	ViT	Multiple new Domains	Parametric Classifier	$k$-Means	DomainNet, SSB-C	$\neq \varnothing$
ArXiv	Wang etal.	ArXiv	ViT	Single New Domain	Parametric Classifier	Known	CIFAR-10, OfficeHome, DomainNet	$\varnothing$

Distribution-Agnostic Category Discovery (DA-CD)

DA-CD eliminates the requirement for a balanced distribution imposed on both labelled and unlabelled data in conventional category discovery. Instead, it acknowledges that the data may follow a skewed distribution, such that for certain categories $`\mathcal{Y}_i`$ and $`\mathcal{Y}_j`$ within the set $`\mathcal{Y}`$ it holds that $`\mathbb{P}_{\mathcal{Y}_x}(\mathcal{Y}_i) > \mathbb{P}_{\mathcal{Y}_x}(\mathcal{Y}_j)`$. In this formulation, the set $`\mathcal{Y}_x`$ may refer to either the labelled categories $`\mathcal{Y}_L`$ or the unlabelled categories $`\mathcal{Y}_U`$.

Year	Method	Pub.	Backbone	Scenario	Label Assignment	# Unlabelled categories	Dataset
2023	NCDLR	TMLR	ViT	Long-tailed Distribution	Parametric Classifier	$k$-Means	CIFAR-10, ImageNet-100, Herb19, iNaturalist18
	ImbaGCD	CVPRW	Resnet	Imbalanced Distribution	Parametric Classifier	Known	CIFAR-10&100, ImageNet-100
	GCDLR	ICCVW	Resnet	Imbalanced Distribution	Parametric Classifier	Known	CIFAR-10&100, ImageNet-100
	BYOP	CVPR	ResNet	Imbalanced Distribution	Parametric Classifier	Known	CIFAR-10&100, TinyImageNet
	BaCon	NeurIPS	ViT	Long-tailed Distribution	$k$-Means	Known	CIFAR-10&100-LT, ImageNet-100-LT, Places-LT
2024	Fan etal	CVPR	ViT	Long-tailed Distribution	$k$-Means	Spectral graph	BioMedical Datasets

Semantic Category Discovery (SCD)

In contrast to NCD and GCD, which focus solely on grouping visually similar images without considering their semantic meaning, SCD extends these paradigms by also assigning a semantic label to each unlabelled instance. Specifically, SCD leverages an open vocabulary label space to achieve this goal. In this context, WordNet, comprising approximately 68,000 labels, is employed as a comprehensive and unconstrained vocabulary, facilitating the assignment of meaningful semantic labels.

Year	Method	Pub.	Backbone	Word Space	Label Assignment	# Unlabelled categories	Dataset
2024	SCD	CVPRW	ViT	~Open	KMeans+Top-k Voting	Known	ImageNet-100&1K, SCars, CUB
	SNCD	AAAI	ResNet	Base + Novel Categories	Parametric Classifier	Known	CIFAR-10&100, ImageNet-100

Few-Shots Category Discovery (FS-CD)

FS-CD addresses the challenge of identifying novel classes when only a very limited amount of labelled data is available. This setting extends traditional category discovery by integrating the principles of few-shot learning. In particular, FSCD adopts an $N$-way, $k$-shot framework in which the model is required to discriminate among $N$ distinct classes with merely $k$ labelled examples per class for base categories.

Chi etal. extend NCD to a few-shot setting by linking it to meta-learning, based on the shared assumption that base and novel categories possess high-level semantic features. By adapting meta-learning techniques such as Model-Agnostic Meta-Learning and Prototypical Networks (ProtoNet), their approach shifts the focus from classification to clustering tasks—a critical adjustment for few-shot category discovery. A key innovation is the introduction of the Clustering-rule-aware Task Sampler, which ensures that training tasks adhere to consistent clustering rules, thereby enabling the model to generalize better to novel categories despite the limited labelled data. However, this method assumes that the number of novel categories is known in advance.

Federated Category Discovery (FCD)

FCD extends Category Discovery in a federated learning setting, facilitating decentralized and collaborative model training among clients while safeguarding data privacy.

Year	Method	Pub.	Backbone	Label Assignment	# Unlabelled categories	Dataset
2023	FedoSSL	ICML	ResNet	Parametric Classifier	Known	CIFAR-10/100, CINIC-10
2024	FedGCD	CVPR	ViT	GMM	Semi-FINCH	CIFAR-10/100, ImageNet-100, CUB, SCars, Pets
Preprint	GAL	ArXiv	ResNet&34	Parametric Classifier	Potential Prototype Merge	CIFAR-100, TinyImageNet, ImageNet-100

Related Repositories

• Visual-AI/Awesome-Category-Discovery — An awesome-style paper list organized by year. Contributions are welcome—feel free to open issues or submit pull requests to add relevant works.
• Awesome-Novel-Class-Discovery — We appreciate this excellent paper collection curated by Joseph, which complements our repo with additional coverage of novel/class discovery works and resources.