CloSeR: Unified Relational Distillation from Closed-Set Teachers for Category Discovery

Yuanpei Liu; Zhenqi He; Jialu Tang; Kai Han

cs.CVAug 26, 2026

CloSeR: Unified Relational Distillation from Closed-Set Teachers for Category Discovery

Authors: Yuanpei Liu, Zhenqi He, Jialu Tang, Kai Han

Organizations: Visual AI Lab, The University of Hong Kong, Hong Kong SAR

Abstract

Generalized Category Discovery (GCD) is an intriguing open-world problem that has garnered increasing attention: given partially labelled data, the goal is to correctly recognize known classes while discovering coherent novel categories from unlabelled samples. Recent GCD methods typically adapt foundation models by jointly optimizing supervised classification and unsupervised discovery objectives on mixed labelled and unlabelled data. While effective, this coupled training can entangle closed-set recognition and open-set discovery, leading to objective conflict and biased predictions, and may disturb the semantic geometry of pretrained representations under limited labels and noisy pseudo-labels. We propose CloSeR, a simple plug-and-play framework that injects Closed-Set Relational knowledge into GCD training. CloSeR first builds a domain-adapted closed-set teacher by tuning lightweight block-wise adapters on labelled known-class data while keeping the foundation model backbone frozen, thereby preserving pretrained priors at low training cost. It then transfers the teacher's knowledge to downstream GCD via Unified Relational Distillation (URD), which distills complementary global sample-to-prototype relations to anchor known-class semantics and local sample-to-sample relations to preserve neighborhood structure, using separate feature pathways to reduce optimization interference. CloSeR is head-agnostic and readily integrates with both parametric and non-parametric GCD methods. Extensive experiments with DINO and DINOv2 backbones on six benchmarks (CIFAR-10/100, ImageNet-100, CUB, Stanford-Cars, and FGVC-Aircraft) show consistent gains over GCD baselines, achieving state-of-the-art performance. Project page: https://visual-ai.github.io/closer/

Figures & tables

Figure 1: The Old -class vs. New -class ACC on CUB [ 58 ] , Stanford-Cars [ 28 ] , and FGVC-Aircraft [ 35 ] using DINOv2 [ 38 ] backbone. We compare two representative GCD baselines (the non-parametric GCD [ 55 ] and the parametric SimGCD [ 63 ] ) with two simple alternatives that first adapt the backbone on labelled known classes (closed-set features) and then perform GCD via either $k$ -means clustering or learning a parametric classifier. Note: “GCD” here denotes the method proposed in [ 55 ] , not the task name.

Figure 2: Overall pipeline of CloSeR (illustrated with SimGCD as an example; it is also applicable to other baselines): (a) Closed-Set Transfer Learning; (b) Unified Relational Distillation. Note : The representation loss in SimGCD is omitted for clarity.

			CIFAR-10 [ 29 ]			CIFAR-100 [ 29 ]			ImageNet-100 [ 11 ]			Average
	Method	[Venue]	All	Old	New	All	Old	New	All	Old	New	All	Old	New
DINO	GCD [ 55 ]	[CVPR 2022]	91.5	97.9	88.2	73.0	76.2	66.5	74.1	89.8	66.3	$79.533\,333\,333\,333\,33$	$87.966\,666\,666\,666\,67$	$73.666\,666\,666\,666\,67$
	PromptCAL [ 70 ]	[CVPR 2023]	97.9	96.6	98.5	81.2	84.2	75.3	83.1	92.7	78.3	$87.4$	$91.166\,666\,666\,666\,67$	$84.033\,333\,333\,333\,33$
	GPC [ 73 ]	[ICCV 2023]	90.6	97.6	87.0	75.4	84.6	60.1	75.3	93.4	66.7	$80.433\,333\,333\,333\,33$	$91.866\,666\,666\,666\,67$	$71.266\,666\,666\,666\,67$
	InfoSieve [ 44 ]	[NeurIPS 2023]	94.8	97.7	93.4	78.3	82.2	70.5	80.5	93.8	73.8	$84.533\,333\,333\,333\,33$	$91.233\,333\,333\,333\,33$	$79.233\,333\,333\,333\,33$
	SPTNet [ 59 ]	[ICLR 2024]	97.3	95.0	98.6	81.3	84.3	75.6	85.4	93.2	81.4	$88$	$90.833\,333\,333\,333\,33$	$85.2$

Table 1: Results on generic datasets (CIFAR-10/100 [ 29 ] and ImageNet-100 [ 11 ] ), reported for All / Old / New categories. Right: average over datasets.

			CUB [ 58 ]			Stanford-Cars [ 28 ]			FGVC-Aircraft [ 35 ]			Average
	Method	[Venue]	All	Old	New	All	Old	New	All	Old	New	All	Old	New
DINO	GCD [ 55 ]	[CVPR 2022]	51.3	56.6	48.7	39.0	57.6	29.9	45.0	41.1	46.9	$45.1$	$51.766\,666\,666\,666\,67$	$41.833\,333\,333\,333\,33$
	PromptCAL [ 70 ]	[CVPR 2023]	62.9	64.4	62.1	50.2	70.1	40.6	52.2	52.2	52.3	$55.1$	$62.233\,333\,333\,333\,33$	$51.666\,666\,666\,666\,67$
	GPC [ 73 ]	[ICCV 2023]	52.0	55.5	47.5	38.2	58.9	27.4	43.3	40.7	44.8	$44.5$	$51.7$	$39.9$
	$\mu$ GCD [ 57 ]	[NeurIPS 2023]	65.7	68.0	64.6	56.5	68.1	50.9	53.8	55.4	53.0	$58.666\,666\,666\,666\,67$	$63.833\,333\,333\,333\,33$	$56.166\,666\,666\,666\,67$
	InfoSieve [ 44 ]	[NeurIPS 2023]	69.4	77.9	65.2	55.7	74.8	46.4	56.3	63.7	52.5	$60.466\,666\,666\,666\,67$	$72.133\,333\,333\,333\,33$	$54.7$

Table 2: Results on fine-grained datasets (CUB [ 58 ] , Stanford-Cars [ 28 ] , FGVC-Aircraft [ 35 ] ), reported for All / Old / New categories. Right: average over datasets.

			CUB [ 58 ]			Stanford-Cars [ 28 ]			FGVC-Aircraft [ 35 ]
Adaptation Method	GCD Method	Pretrained	All	Old	New	All	Old	New	All	Old	New
(1) Full Fine-tuning	SimGCD	DINO	68.1	76.3	64.0	60.9	77.5	52.9	54.8	53.2	55.6
(2) Fine-tuning Last Block	SimGCD	DINO	58.5	56.2	59.6	46.7	56.8	40.3	45.8	41.8	47.8
(3) Block-wise Adapter	SimGCD	DINO	68.2	76.9	63.9	63.0	80.2	54.7	56.6	55.5	57.2
(1) Full Fine-tuning	SimGCD	DINOv2	76.9	77.4	76.6	78.4	91.7	72.0	67.9	62.6	70.5
(2) Fine-tuning Last Block	SimGCD	DINOv2	77.3	80.2	75.8	77.9	90.5	71.8	67.7	64.6	69.2

Table 3: Experimental results using different transfer learning methods with DINO [ 4 ] and DINOv2 [ 38 ] pretrained weights on SimGCD. Results are reported on SSB [ 56 ] .

Global Relation	Local Relation	Feature Decoupling	All	Old	New	All	Old	New	All	Old	New
			CUB [ 58 ]			Stanford-Cars [ 28 ]			FGVC-Aircraft [ 35 ]
✗	✗	✗	73.6	75.3	72.8	58.5	75.6	50.3	57.1	64.7	53.3
✓	✗	✗	80.5	78.9	81.2	61.9	81.0	52.6	59.0	65.1	56.0
✗	✓	✗	80.2	78.5	81.1	64.3	82.3	55.7	63.0	66.7	61.1
✓	✓	✗	81.0	78.5	82.2	65.8	82.8	57.6	64.9	68.4	63.2
✓	✓	✓	82.8	80.3	84.1	66.5	83.4	58.3	66.6	68.1	65.8

Table 4: Experimental results using different distillation components with DINO [ 4 ] pretrained weights on SelEx. Results are reported on the SSB [ 56 ] benchmark.

			CUB [ 58 ]			Stanford-Cars [ 28 ]			FGVC-Aircraft [ 35 ]			CIFAR-10 [ 29 ]			CIFAR-100 [ 29 ]			ImageNet-100 [ 11 ]
Method	Teacher	Student	All	Old	New	All	Old	New	All	Old	New	All	Old	New	All	Old	New	All	Old	New
SimGCD	-	ViT-B	71.5	78.1	68.3	71.5	81.9	66.6	63.9	69.9	60.9	98.7	96.7	99.7	88.5	89.2	87.2	89.9	95.5	87.1
+ CloSeR	ViT-B	ViT-B	78.1	79.5	77.3	78.8	90.7	73.0	71.8	68.7	73.3	98.9	97.1	99.8	93.2	93.0	93.5	92.5	96.6	90.4
+ CloSeR	ViT-L	ViT-B	81.3	81.4	81.3	79.9	92.5	73.8	75.0	77.7	73.6	99.4	98.9	99.7	93.8	94.2	93.0	93.4	95.8	92.2
SimGCD	-	ViT-L	71.0	73.7	69.6	71.3	80.8	66.7	70.6	70.1	70.8	99.5	98.7	99.8	91.8	91.8	91.8	92.4	94.0	91.7
+ CloSeR	ViT-L	ViT-L	81.4	83.2	80.5	83.8	92.9	79.4	76.2	82.4	73.0	99.5	98.7	99.9	94.5	94.9	93.6	93.7	97.2	91.9

Table 5: Experimental results using different backbone sizes with DINOv2 [ 38 ] on SimGCD and SimGCD+CloSeR. Results are reported on SSB [ 56 ] .

Figure 3: Comparison of attention maps from the backbone networks of SimGCD [ 63 ] and our SimGCD+CloSeR. The results include three datasets: Stanford-Cars [ 28 ] , CUB [ 58 ] , and FGVC-Aircraft [ 35 ] , spanning both Old and New classes.

Generalized Category Discovery

Generalized category discovery learns known classes from labeled examples while grouping unlabeled data into known and previously unseen categories.

See all 25 papers

Sep 9, 2026

CoGe-GCD: Reframing Generalized Category Discovery with Compositional Generalization

Generalized Category Discovery (GCD) assigns unlabeled instances, mixed with labeled data, to known or novel categories, requiring human-like compositional reasoning: reusing primitives learned from known classes and deciding when new combinations imply new categories. Existing GCD methods operate on unstructured token features and struggle to extrapolate to novel compositions. We propose CoGe-GCD, which rethinks GCD through compositional generalization with two coupled stages. (i) Compositional Perception structures patch tokens by mapping them to a small vocabulary of primitives and refining token embeddings via competitive token-primitive assignment and information passing, yielding coherent groups for discovery. (ii) Generalizing Induction exploits the induced geometric structure and applies a structure-preserving calibration over spatial relations, maintaining probabilistic semantics while improving extrapolation to unseen primitive combinations. CoGe-GCD is implemented as an inductive-bias module between backbone and projection head, without modifying heads or losses, and can be plugged into diverse GCD frameworks. On standard benchmarks, it consistently improves all-class accuracy, unknown-class number estimation, and geometric quality, with marginal computational overhead. Code is available at https://github.com/lytang63/CoGe-GCD.

Luyao Tang, Jiewei Zheng, Kunze Huang +3

The University of Hong Kong, Hong Kong SAR, China · Xiamen University, Xiamen, China · Shenzhen University, Shenzhen, China.

Geometric Representation Learning Classification Compositional Generalization Generalized Category Discovery

Ask PDF

May 10, 2026

Relational Retrieval: Leveraging Known-Novel Interactions for Generalized Category Discovery

In this study, we tackle Generalized Category Discovery (GCD) via a Relational Retrieval perspective, explicitly coupling labeled and unlabeled data through bidirectional knowledge transfer. While existing methods treat these sources separately, missing valuable interaction opportunities, we propose Relational Pattern Consistency (RPC) that enables mutual enhancement. RPC employs One-vs-All classifiers for soft ID/OOD decomposition, then introduces two mechanisms: (i) for known-class preservation, we transfer semantic behavioral alignment; (ii) for category discovery, we leverage the insight that samples from the same category maintain invariant relationships with known-class prototypes, transforming unreliable pseudo-labeling into well-defined relational pattern matching. This bidirectional design allows labeled data to guide unlabeled learning while discovering novel categories through their collective relational signatures. Extensive experiments demonstrate RPC achieves state-of-the-art performance on both generic and fine-grained benchmarks.

Yulin Xu, Chunqi Guo, Yuanzhen Shuai +1

University College London London, UK · Juniata College Huntingdon, USA

Semi-Supervised Learning Generalized Category Discovery

Ask PDF

Apr 16, 2026

OmniGCD: Abstracting Generalized Category Discovery for Modality Agnosticism

Generalized Category Discovery (GCD) challenges methods to identify known and novel classes using partially labeled data, mirroring human category learning. Unlike prior GCD methods, which operate within a single modality and require dataset-specific fine-tuning, we propose a modality-agnostic GCD approach inspired by the human brain's abstract category formation. Our

\textbf{OmniGCD}

leverages modality-specific encoders (e.g., vision, audio, text, remote sensing) to process inputs, followed by dimension reduction to construct a

\textbf{GCD latent space}

, which is transformed at test-time into a representation better suited for clustering using a novel synthetically trained Transformer-based model. To evaluate OmniGCD, we introduce a

\textbf{zero-shot GCD setting}

where no dataset-specific fine-tuning is allowed, enabling modality-agnostic category discovery.

\textbf{Trained once on synthetic data}

, OmniGCD performs zero-shot GCD across 16 datasets spanning four modalities, improving classification accuracy for known and novel classes over baselines (average percentage point improvement of

\textbf{+6.2}

\textbf{+17.9}

\textbf{+1.5}

and

\textbf{+12.7}

for vision, text, audio and remote sensing). This highlights the importance of strong encoders while decoupling representation learning from category discovery. Improving modality-agnostic methods will propagate across modalities, enabling encoder development independent of GCD. Our work serves as a benchmark for future modality-agnostic GCD works, paving the way for scalable, human-inspired category discovery. All code is available here

Jordan Shipard, Arnold Wiliem, Kien Nguyen Thanh +2

SAIVT, QUT, Australia · Shield AI, Australia · La Trobe University, Australia

Generalized Category Discovery

Ask PDF

Abstract

Generalized Category Discovery

Explore similar work

CoGe-GCD: Reframing Generalized Category Discovery with Compositional Generalization

Relational Retrieval: Leveraging Known-Novel Interactions for Generalized Category Discovery

OmniGCD: Abstracting Generalized Category Discovery for Modality Agnosticism