Generalized Category Discovery (GCD) aims to perform old-category recognition and novel-category discovery in the same unlabeled set. Early methods address both objectives through clustering and evaluate the exchangeable predicted indices with global Hungarian (GH) matching. Recent methods adopt trainable parametric heads, using supervision for old-category recognition and discovery-oriented objectives for new categories. In this way, the output space becomes partially exchangeable: old indices are semantically anchored, while new indices remain exchangeable. However, these methods still apply global matching, which can reassign anchored old indices to other categories and hide some prediction errors. Empirically, we are surprised to observe that this reassignment can induce an overestimation of new-category accuracy by up to 25.63 points. We term this phenomenon semantic mismapping and further derive a permutation-cycle condition that characterizes when semantic mismapping can increase the global matching score. To fix this issue, we propose Semantics-Preserving Evaluation (SPE), which preserves the supervised semantics of old indices and restricts Hungarian matching to the new region. Furthermore, we propose Mismapping-Mitigating Optimization (MMO) to alleviate the prediction errors that can give rise to semantic mismapping, combining a region-balance loss (RBL) for old/new routing with a novel-diversity loss (NDL) that encourages orthogonality among new prototypes. Theoretically, we show that minimizing the two losses can tighten an upper bound on spurious accuracy gains from semantic mismapping, with controllable residual terms. Empirical results under both GH and SPE demonstrate the effectiveness of our proposed method.
Introduction
Traditional recognition tasks typically follow a closed-world assumption: the categories encountered at test time are covered by the labeled training set. However, in real-world scenarios, a deployed recognition system may receive samples from known categories together with samples from categories that have never been labeled before [1, 2]. In view of this, Generalized Category Discovery (GCD) considers a more challenging open-world setting, where a model is trained with labeled samples from old categories and unlabeled samples drawn from both old and new categories [3]. This setting gives rise to two objectives that typically belong to different learning paradigms: (O1) recognizing old categories in the unlabeled samples (i.e., classification) and (O2) discovering underlying clusters for new categories (i.e., clustering). Hence, it becomes a core question in GCD: how to unify the two distinct objectives into a single framework?
Early GCD methods adopt a unified clustering-based framework, learning visual features from both labeled and unlabeled data and then applying non-parametric clustering [3, 4, 5]. The induced cluster indices are exchangeable as their semantics have been lost. Hence, the standard GCD evaluation protocol aligns these indices with ground-truth categories through Hungarian matching [6] and then calculates accuracy.
Recent advances, represented by SimGCD [7] and its follow-ups [8, 9, 10], turn to a divide-and-conquer parametric framework, achieving superior performance. These methods replace the non-parametric clustering procedure with trainable classifier or prototype heads. For old categories (O1), the parametric head is effectively trained with the supervision from labeled samples. For new categories (O2), the corresponding indices are not tied to fixed categories and are trained with discovery-oriented objectives, such as self-distillation [7], pseudo-labeling [8], and contrastive learning [11]. Overall, the output space becomes partially exchangeable: old indices are semantically anchored, while new indices remain exchangeable.
Although effective, this divide-and-conquer framework can exhibit notorious prediction bias. Since old categories receive more direct supervision, the parametric head tends to produce an imbalanced distribution across old and new categories [7]. A typical failure is false-old, where many new samples are predicted as old categories. To address this bias, prior work has designed a series of training objectives and regularizers [7, 9, 12, 13].
Despite their progress, parametric methods still apply Hungarian matching to the entire output space. However, we find that such global matching can reassign semantically anchored old indices to new categories, making some false-old predictions appear correct. As illustrated in Figure 1(a), a new bird with an incorrect old-category prediction is reassigned to its ground-truth new category. In view of this, we call any assignment that overrides an anchored index semantic mismapping, which may occur both across the old/new boundary and within the old region. Theoretically, we further derive a permutation-cycle condition that characterizes when semantic mismapping can increase the global matching score. For fine-grained datasets, such as Herbarium 19, this mismapping can induce an overestimation of new accuracy by 25.63 points, whereas for coarse-grained datasets, such as CIFAR-10 and ImageNet-100, its effect becomes negligible. This observation raises a natural question:
Figure 1. (a) When evaluating parametric GCD methods, global matching could reassign an old output to a new category (e.g., o1→n1), making a false-old prediction appear correct. (b) Our SPE protocol preserves the semantics of old outputs and performs local matching within the new region. (c) Our mismapping-mitigating optimization (MMO) combines region-balance loss (RBL), which matches old/new prediction mass to estimated sample proportions, with novel-diversity loss (NDL), which encourages orthogonality among new prototypes.
How should parametric GCD methods be evaluated to avoid semantic mismapping?
To answer this question, we propose a divide-and-conquer protocol named Semantics-Preserving Evaluation (SPE). As illustrated in Figure 1(b), SPE preserves the semantics of old outputs and performs local matching within the new region, while predictions crossing the old/new boundary remain errors. This design directly measures old-category recognition for (O1) while preserving the permutation invariance required for novel-category discovery under (O2). We further extend SPE to non-parametric GCD methods and settings with an estimated number of new categories. Motivated by our analysis, we propose Mismapping-Mitigating Optimization (MMO), which combines two plug-and-play regularizers following the divide-and-conquer principle, as summarized in Figure 1(c). On the one hand, a region-balance loss (RBL) explicitly balances prediction mass across the old and new regions. On the other hand, a novel-diversity loss (NDL) promotes new-category discrimination by encouraging orthogonality among new prototypes. We further derive two bounds showing how the two losses can suppress spurious accuracy gains when residual errors are controlled. These bounds provide sufficient conditions for GH and SPE to yield the same overall accuracy. Finally, extensive experiments on several challenging fine-grained benchmarks validate the necessity of SPE and the effectiveness of the proposed regularizers. Overall, our contributions are summarized as follows:
New insight. We identify semantic mismapping in parametric GCD and theoretically characterize when it produces spurious accuracy gains under GH.
New protocol. Building on this insight, we propose SPE to prevent semantic mismapping from hiding prediction errors, enabling faithful evaluation of old-category recognition and novel-category discovery.
New method. To mitigate the underlying prediction errors, we propose MMO with two plug-and-play losses, supported by theoretical bounds and extensive experiments.
Related Work
Category-Discovery Settings
Category discovery aims to identify previously unseen categories in unlabeled data by transferring knowledge from labeled old categories [14, 15, 16]. As one of the most representative settings, Novel Category Discovery (NCD) provides labeled old-category data alongside an unlabeled discovery set containing only new categories [17, 18, 19, 20, 21]. Mainstream NCD methods combine self-supervised learning with old-category supervision to improve the clustering of unlabeled samples [18, 22, 23, 24, 25]. Class-incremental NCD extends this setting by first learning on labeled old-category data and then receiving unlabeled data for novel-category discovery [26, 27]. Generalized Category Discovery (GCD) is a more challenging setting, where old and new categories coexist in the unlabeled set [3, 28]. Subsequent studies investigate fine-grained discovery [29, 30, 31], domain shifts [32], and continuous discovery [33]. OpenGCD integrates generalized category discovery into open-world recognition [34, 35] to assist in grouping unlabeled samples for subsequent incremental learning [36].
This paper focuses on the standard GCD setting, studying how to faithfully evaluate and jointly improve old-category recognition and novel-category discovery.
Evaluation Protocols for Category Discovery
Evaluation protocols for the aforementioned settings mainly differ in how they emphasize old-category recognition and novel-category discovery. Early NCD studies focus on clustering accuracy for new categories [17, 18]. Later NCD studies evaluate both old and new categories, distinguishing task-aware and task-agnostic protocols according to whether the model is told if each sample belongs to an old or a new category [37, 38, 39]. Class-incremental NCD also measures how well old-category recognition is maintained as the model learns new categories [26, 27]. Notably, FRoST also identifies old–new errors hidden by Hungarian matching and preserves old-category semantics during NCD evaluation [27]. However, its new-category mapping is established on new-only training data, leaving evaluation on the mixed unlabeled set in GCD unaddressed.
To handle the coexistence of old and new categories, GCD extends NCD evaluation by applying a global Hungarian matching to the whole unlabeled set [3]. This clustering-based protocol remains widely used by parametric GCD methods, providing a common basis for comparison with non-parametric approaches [7, 12]. μGCD also observes that Hungarian matching can report nonzero new-category accuracy even when no samples are predicted to new outputs [8]. However, its evaluation still follows global matching, without preserving anchored old-output semantics. We systematically study this mismatch in standard GCD, characterize matching gains through permutation cycles, quantify their effects on empirical performance and method rankings, and propose SPE for faithful evaluation.
Recent work also highlights the distinction between clustering accuracy and semantic classification. SECOS uses category names and vision-language knowledge to classify both old and new categories, without post-hoc matching [40]. In standard GCD, however, new categories are expected to be discovered without their names being given. Hence, our work focuses on evaluation in the standard GCD setting.
GCD Methods and Prediction Bias
Early GCD approaches learn transferable features from both labeled and unlabeled data, followed by non-parametric clustering [3, 4, 5]. Further studies improve feature adaptation through spatial prompts or textual knowledge [10, 11, 42]. Parametric methods replace non-parametric clustering with trainable classifiers or prototypes [7, 8, 12]. This divide-and-conquer paradigm achieves state-of-the-art performance, as the parametric classifier leverages old-category supervision more effectively.
However, this parametric paradigm can also lead to prediction bias, as old and new categories rely on different learning signals. Direct supervision can bias predictions toward old categories [7], while overconfidence in new categories can also hinder old-category recognition accuracy [43]. To mitigate this bias, existing methods apply constraints at three levels. At the prediction level, existing methods use entropy and class-wise regularization [7, 13], consistency constraints [9], or auxiliary debiased classifiers [43, 44] to balance old- and new-category predictions. At the feature level, existing methods use prototype separation [12], feature-geometry regularization [45], or relational distillation [46] to improve category discrimination. At the gradient level, EAGC coordinates supervised and unsupervised gradients to reduce conflicts between old-category recognition and novel-category discovery [47].
Our work instead examines prediction bias at the evaluation level, revealing how global matching can hide the resulting errors. Guided by this insight, we design regularizers at the prediction and feature levels.
Preliminaries
Generalized Category Discovery
Let Dl={(xi,yi)}i=1Nl and Du={xj}j=1Nu denote the labeled and unlabeled training sets, respectively. The labels in Dl belong to a set of old categories Yold, whereas Du contains samples from both Yold and a set of new categories Ynew. The two category sets are disjoint, and together form the target space Y=Yold∪Ynew. Given Dl and Du, GCD aims to learn a model f:X→Y^, where Y^ denotes the predicted output indices. Unless otherwise specified, we assume that the total number of target categories is given and set ∣Y^∣=∣Y∣, following the common practice [3, 7]. When this number is unknown, ∣Y^∣ can be estimated via off-the-shelf methods [3, 48]. During evaluation, f is applied to each sample in Du, recognizing old-category samples and assigning new-category samples to proper category indices.
Early GCD methods follow a unified clustering-based framework[3, 4, 5]. Specifically, the model can be formulated as f=ψ∘ϕ, where ϕ:X→Rd is a feature extractor and ψ:Rd→Y^ is a non-parametric clustering procedure. In this framework, Y^ is fully exchangeable and will be aligned with the ground-truth categories during evaluation.
Recent methods instead adopt a divide-and-conquer parametric framework[7, 8, 9, 10]. These methods replace ψ with a trainable classifier or prototype head h:Rd→Y^, so that the model can be formulated as f=h∘ϕ. In this case, the output space is partitioned as Y^=Y^old∪Y^new and is only partially exchangeable: the old-category dimensions in Y^old are semantically anchored by the supervision from Dl, while the new-category dimensions in Y^new remain exchangeable and are learned through discovery-oriented objectives defined on Du.
Standard GCD Evaluation Protocol
The standard GCD protocol treats every predicted index as exchangeable and finds one assignment over Du with the Hungarian algorithm [3, 6]. Let Π be the set of one-to-one assignments from Y^ to Y. The evaluator has access to the ground-truth category y(x) and obtains
AGH∈argA∈Πmaxx∈Du∑I{A(f(x))=y(x)}.(1)
This single assignment is then reused for every reported subset. For r∈{old,new,all}, define Du,r:={x∈Du:y(x)∈Yr}, where Yall=Y. The corresponding accuracy is
For conciseness, we write these three scores as OldGH, NewGH, and AllGH, respectively.
This protocol is appropriate when all output indices are exchangeable. However, a parametric GCD head is only partially exchangeable. This mismatch can induce a problem we call semantic mismapping, which we will analyze in the next section.
Semantic Mismapping and Semantics-Preserving Evaluation
Semantic Mismapping under Global Matching
To illustrate semantic mismapping, we first partition the confusion matrix into four blocks, with predicted output groups as rows and ground-truth category groups as columns:
Yold
Ynew
Y^old
Yold×Y^old
Ynew×Y^old
Y^new
Yold×Y^new
Ynew×Y^new
The diagonal blocks, i.e., Yold×Y^old and Ynew×Y^new, represent old-to-old and new-to-new predictions, respectively.
In contrast, the off-diagonal blocks, i.e., Yold×Y^new and Ynew×Y^old, represent old-to-new and new-to-old errors, respectively.
How does global matching induce semantic mismapping? Next, we take the predictions in Figure 2 as an example. In this case, keeping o^1↦o1 and n^1↦n1 only yields 7+5=12 correct predictions, whereas swapping the two mappings yields 46+44=90 seemingly correct ones. Thus, global matching will choose A(o^1)=n1 and A(n^1)=o1, counting both groups of cross-boundary errors as correct, as shown in Figure 2(b). Inspired by this example, we have the following definition of semantic mismapping.
Figure 2. Comparison of GH and SPE using the same predictions. (a) The raw confusion matrix has predicted outputs as rows and ground-truth categories as columns, with old and new categories denoted by o and n. (b) GH reassigns the anchored old output o^1 to n1 and the new output n^1 to o1, counting 46 false-old predictions and 44 old-to-new errors as correct. (c) SPE preserves the semantics of old outputs and applies local matching within the new region. In this way, predictions in either cross-region block remain errors, including the 46 and 44 samples hidden by GH.
Definition 1 (Semantic mismapping). For each old category yo∈Yold, let y^o∈Y^old denote its output index in the parametric head h. Given an assignment A, semantic mismapping occurs at y^o if A(y^o)=yo.
References
[1]D. Hendrycks and K. Gimpel, “A baseline for detecting misclassified and out-of-distribution examples in neural networks,” in International Conference on Learning Representations, 2017, pp. 1–12.
[2]Z. Xu, Q. Xu, Z. Wang, C. Hua, S. Li, Z. Yang, and Q. Huang, “Mind the way you select negative texts: Pursuing the distance consistency in OOD detection with VLMs,” in IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2026, pp. 34 650–34 661.arxiv.org/abs/2603.02618
[3]S. Vaze, K. Han, A. Vedaldi, and A. Zisserman, “Generalized category discovery,” in IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 7482–7491.arxiv.org/abs/2201.02609
[4]N. Pu, Z. Zhong, and N. Sebe, “Dynamic conceptional contrastive learning for generalized category discovery,” in IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 7579–7588.arxiv.org/abs/2303.17393
[5]S. Zhang, S. Khan, Z. Shen, M. Naseer, G. Chen, and F. S. Khan, “PromptCAL: Contrastive affinity learning via auxiliary prompts for generalized novel category discovery,” in IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 3479–3488.arxiv.org/abs/2212.05590
[6]H. W. Kuhn, “The Hungarian method for the assignment problem,” Naval Research Logistics Quarterly, vol. 2, no. 1–2, pp. 83–97, 1955.
[7]X. Wen, B. Zhao, and X. Qi, “Parametric classification for generalized category discovery: A baseline study,” in IEEE/CVF International Conference on Computer Vision, 2023, pp. 16 544–16 554.arxiv.org/abs/2211.11727
[8]S. Vaze, A. Vedaldi, and A. Zisserman, “No representation rules them all in category discovery,” in Advances in Neural Information Processing Systems, 2023, pp. 19 962–19 989.arxiv.org/abs/2311.17055
[9]X. Cao, X. Zheng, G. Wang, W. Yu, Y. Shen, K. Li, Y. Lu, and Y. Tian, “Solving the catastrophic forgetting problem in generalized category discovery,” in IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 16 880–16 889.arxiv.org/abs/2501.05272
[10]E. Wang, Z. Peng, Z. Xie, F. Yang, X. Liu, and M. Cheng, “GET: unlocking the multi-modal potential of CLIP for generalized category discovery,” in IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2025, pp. 20 296–20 306.arxiv.org/abs/2403.09974
[11]S. Choi, D. Kang, and M. Cho, “Contrastive mean-shift learning for generalized category discovery,” in IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 23 094–23 104.arxiv.org/abs/2404.09451
[12]S. Ma, F. Zhu, X.-Y. Zhang, and C.-L. Liu, “ProtoGCD: Unified and unbiased prototype learning for generalized category discovery,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 47, no. 7, pp. 6022–6038, 2025.DOI
[13]D. Liu, Z. Tan, L. Zhao, Z. Zhang, X. Fang, and W. Huang, “Generalized category discovery via reciprocal learning and class-wise distribution regularization,” in International Conference on Machine Learning, 2025, pp. 39 651–39 668.arxiv.org/abs/2506.02334
[14]Y. Hsu, Z. Lv, and Z. Kira, “Learning to cluster in order to transfer across domains and tasks,” in International Conference on Learning Representations, 2018, pp. 1–20.arxiv.org/abs/1711.10125
[15]Y. Hsu, Z. Lv, J. Schlosser, P. Odom, and Z. Kira, “Multi-class classification without multi-class labels,” in International Conference on Learning Representations, 2019, pp. 1–16.arxiv.org/abs/1901.00544
[16]Z. Zhong, L. Zhu, Z. Luo, S. Li, Y. Yang, and N. Sebe, “OpenMix: Reviving known knowledge for discovering novel visual categories in an open world,” in IEEE Conference on Computer Vision and Pattern Recognition, 2021, pp. 9462–9470.arxiv.org/abs/2004.05551
[17]K. Han, A. Vedaldi, and A. Zisserman, “Learning to discover novel visual categories via deep transfer clustering,” in IEEE/CVF International Conference on Computer Vision, 2019, pp. 8400–8408.arxiv.org/abs/1908.09884
[18]K. Han, S. Rebuffi, S. Ehrhardt, A. Vedaldi, and A. Zisserman, “Automatically discovering and learning new visual categories with ranking statistics,” in International Conference on Learning Representations, 2020, pp. 1–13.arxiv.org/abs/2002.05714
[19]H. Chi, F. Liu, W. Yang, L. Lan, T. Liu, B. Han, G. Niu, M. Zhou, and M. Sugiyama, “Meta discovery: Learning to discover novel classes given very limited data,” in International Conference on Learning Representations, 2022, pp. 1–20.arxiv.org/abs/2102.04002
[20]M. Yang, L. Wang, C. Deng, and H. Zhang, “Bootstrap your own prior: Towards distribution-agnostic novel class discovery,” in IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 3459–3468.DOI
[21]Y. Liu, Y. Cai, Q. Jia, B. Qiu, W. Wang, and N. Pu, “Novel class discovery for ultra-fine-grained visual categorization,” in IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 17 679–17 688.arxiv.org/abs/2405.06283
[22]Z. Zhong, E. Fini, S. Roy, Z. Luo, E. Ricci, and N. Sebe, “Neighborhood contrastive learning for novel class discovery,” in IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021, pp. 10 867–10 875.arxiv.org/abs/2106.10731
[23]X. Jia, K. Han, Y. Zhu, and B. Green, “Joint representation learning and novel category discovery on single- and multi-modal data,” in IEEE/CVF International Conference on Computer Vision, 2021, pp. 590–599.arxiv.org/abs/2104.12673
[24]K. Han, S. Rebuffi, S. Ehrhardt, A. Vedaldi, and A. Zisserman, “AutoNovel: Automatically discovering and learning novel visual categories,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 44, no. 10, pp. 6767–6781, 2022.DOI
[25]B. Zhao and K. Han, “Novel visual category discovery with dual ranking statistics and mutual knowledge distillation,” in Advances in Neural Information Processing Systems, 2021, pp. 22 982–22 994.arxiv.org/abs/2107.03358
[26]K. I. Joseph, S. Paul, K. Aggarwal, S. Biswas, P. Rai, K. Han, and V. N. Balasubramanian, “Novel class discovery without forgetting,” in European Conference on Computer Vision, 2022, pp. 570–586.arxiv.org/abs/2207.10659
[27]S. Roy, M. Liu, Z. Zhong, N. Sebe, and E. Ricci, “Class-incremental novel class discovery,” in European Conference on Computer Vision, 2022, pp. 317–333.
[28]R. Ouldnough, C. Kuo, and Z. Kira, “CLIP-GCD: simple language guided generalized category discovery,” CoRR, vol. abs/2305.10420, 2023.
[29]Y. Fei, Z. Zhao, S. Yang, and B. Zhao, “XCon: Learning with experts for fine-grained category discovery,” in British Machine Vision Conference, 2022, pp. 1–12.arxiv.org/abs/2208.01898
[30]S. Rastegar, M. Salehi, Y. M. Asano, H. Doughty, and C. G. M. Snoek, “SelEx: Self-expertise in fine-grained generalized category discovery,” in European Conference on Computer Vision, 2024, pp. 440–458.
[31]H. Zheng, N. Pu, W. Li, N. Sebe, and Z. Zhong, “Generalized fine-grained category discovery with multi-granularity conceptual experts,” Int. J. Comput. Vis., vol. 134, no. 8, p. 379, 2026.
[32]H. Wang, S. Vaze, and K. Han, “HiLo: A learning framework for generalized category discovery robust to domain shifts,” in International Conference on Learning Representations, 2025, pp. 1–42.arxiv.org/abs/2408.04591
[33]H. Kim, S. Suh, D. Kim, D. Jeong, H. Cho, and J. Kim, “Proxy anchor-based unsupervised learning for continuous generalized category discovery,” in IEEE/CVF International Conference on Computer Vision, 2023, pp. 16 642–16 651.arxiv.org/abs/2307.10943
[34]Z. Wang, Q. Xu, Z. Yang, Y. He, X. Cao, and Q. Huang, “OpenAUC: Towards AUC-oriented open-set recognition,” in Advances in Neural Information Processing Systems, 2022, pp. 25 033–25 045.arxiv.org/abs/2210.13458
[35]C. Hua, Q. Xu, Z. Yang, Z. Wang, S. Bao, and Q. Huang, “OpenworldAUC: Towards unified evaluation and optimization for open-world prompt tuning,” in International Conference on Machine Learning, 2025, pp. 24 975–25 020.arxiv.org/abs/2505.05180
[36]F. Gao, W. Zhong, Z. Cao, X. Peng, and Z. Li, “OpenGCD: Assisting open world recognition with generalized category discovery,” CoRR, vol. abs/2308.06926, 2023.arxiv.org/abs/2308.06926
[37]E. Fini, E. Sangineto, S. Lathuilière, Z. Zhong, M. Nabi, and E. Ricci, “A unified objective for novel class discovery,” in IEEE/CVF International Conference on Computer Vision, 2021, pp. 9264–9272.arxiv.org/abs/2108.08536
[38]M. Yang, Y. Zhu, J. Yu, A. Wu, and C. Deng, “Divide and conquer: Compositional experts for generalized novel class discovery,” in IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 14 248–14 257.DOI
[39]W. Li, Z. Fan, J. Huo, and Y. Gao, “Modeling inter-class and intra-class constraints in novel class discovery,” in IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 3449–3458.arxiv.org/abs/2210.03591
[40]H. Liu, J. Yang, J. Gao, M. Li, Y. Zhang, S. N. Gowda, and Y. Lu, “SECOS: Semantic capture for rigorous classification in open-world semi-supervised learning,” in IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2026, pp. 39 627–39 636.
[42]H. Zheng, N. Pu, W. Li, N. Sebe, and Z. Zhong, “Textual knowledge matters: Cross-modality co-teaching for generalized visual class discovery,” in European Conference on Computer Vision, 2024, pp. 41–58.arxiv.org/abs/2403.07369
[43]P. Guo, Y. Song, and B. Wang, “Towards debiased generalized category discovery,” in International Joint Conference on Artificial Intelligence, 2025, pp. 5271–5279.DOI
[44]Y. Liu and K. Han, “DebGCD: Debiased learning with distribution guidance for generalized category discovery,” in International Conference on Learning Representations, 2025, pp. 1–14.arxiv.org/abs/2504.04804
[45]L. Tang, K. Huang, C. Chen, and C. Chen, “Bures-isotropy alignment: Manifold learning of generalized category discovery,” in International Conference on Learning Representations, 2026, pp. 1–25.
[46]Y. Liu, Z. He, J. Tang, and K. Han, “CloSeR: Unified relational distillation from closed-set teachers for category discovery,” in European Conference on Computer Vision, 2026, pp. 1–18.arxiv.org/abs/2608.25692
[47]H. Zheng, N. Pu, Y. Cai, T. Long, W. Li, N. Sebe, and Z. Zhong, “The devil is in gradient entanglement: Energy-aware gradient coordinator for robust generalized category discovery,” in IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2026, pp. 3563–3573.arxiv.org/abs/2604.14176
[48]B. Zhao, X. Wen, and K. Han, “Learning semi-supervised gaussian mixture models for generalized category discovery,” in IEEE/CVF International Conference on Computer Vision, 2023, pp. 16577–16587.arxiv.org/abs/2305.06144
[49]J. Bekker and J. Davis, “Estimating the class prior in positive and unlabeled data through decision tree induction,” in AAAI Conference on Artificial Intelligence, 2018, pp. 2712–2719.DOI
[50]A. Krizhevsky and G. Hinton, “Learning multiple layers of features from tiny images,” University of Toronto, Technical Report, 2009.
[51]J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei, “ImageNet: A large-scale hierarchical image database,” in IEEE Conference on Computer Vision and Pattern Recognition, 2009, pp. 248–255.
[52]O. Russakovsky, J. Deng, H. Su, J. Krause, S. Satheesh, S. Ma, Z. Huang, A. Karpathy, A. Khosla, M. S. Bernstein, A. C. Berg, and L. Fei-Fei, “ImageNet large scale visual recognition challenge,” Int. J. Comput. Vis., vol. 115, no. 3, pp. 211–252, 2015.
[53]C. Wah, S. Branson, P. Welinder, P. Perona, and S. Belongie, “The Caltech-UCSD Birds-200-2011 dataset,” California Institute of Technology, Technical Report CNS-TR-2011-001, 2011.
[54]J. Krause, M. Stark, J. Deng, and L. Fei-Fei, “3D object representations for fine-grained categorization,” in IEEE International Conference on Computer Vision Workshops, 2013, pp. 554–561.DOI
[55]S. Maji, E. Rahtu, J. Kannala, M. B. Blaschko, and A. Vedaldi, “Fine-grained visual classification of aircraft,” CoRR, vol. abs/1306.5151, 2013.arxiv.org/abs/1306.5151
[56]K. C. Tan, Y. Liu, B. Ambrose, M. Tulig, and S. J. Belongie, “The herbarium challenge 2019 dataset,” CoRR, vol. abs/1906.05372, 2019.arxiv.org/abs/1906.05372
[57]M. Caron, H. Touvron, I. Misra, H. Jégou, J. Mairal, P. Bojanowski, and A. Joulin, “Emerging properties in self-supervised vision transformers,” in IEEE/CVF International Conference on Computer Vision, 2021, pp. 9630–9640.arxiv.org/abs/2104.14294
[58]A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark, G. Krueger, and I. Sutskever, “Learning transferable visual models from natural language supervision,” in International Conference on Machine Learning, 2021, pp. 8748–8763.arxiv.org/abs/2103.00020
[59]X. Yang, Y. Zhang, S. Mi, and X.-S. Wei, “TAR: Token- aware refinement for fine-grained generalized category discovery,” in IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2026, pp. 31 995–32 004.
[60]A. Paszke, S. Gross, F. Massa, A. Lerer, J. Bradbury, G. Chanan, T. Killeen, Z. Lin, N. Gimelshein, L. Antiga, A. Desmaison, A. Köpf, E. Z. Yang, Z. DeVito, M. Rai- son, A. Tejani, S. Chilamkurthy, B. Steiner, L. Fang, J. Bai, and S. Chintala, “PyTorch: An imperative style, high-performance deep learning library,” in Advances in Neural Information Processing Systems, 2019, pp. 8024– 8035.arxiv.org/abs/1912.01703
[61]A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weis- senborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gelly, J. Uszkoreit, and N. Houlsby, “An image is worth 16x16 words: Trans- formers for image recognition at scale,” in International Conference on Learning Representations, 2021, pp. 1–22.
[62]Z. Wang, Q. Xu, Z. Yang, Y. He, X. Cao, and Q. Huang, “A unified generalization analysis of re-weighting and logit-adjustment for imbalanced learning,” in Advances in Neural Information Processing Systems, 2023, pp. 48 417–48 430.DOI
[63]Z. Wang, Q. Xu, Z. Yang, Z. Xu, L. Zhang, X. Cao, and Q. Huang, “A unified perspective for loss-oriented im- balanced learning via localization,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 48, no. 1, pp. 639–656, 2026.DOI
[64]Z. Yang, Q. Xu, S. Li, Z. Wang, X. Cao, and Q. Huang, “DirMixE: Harnessing test agnostic long-tail recogni- tion with hierarchical label variations,” IEEE Transac- tions on Pattern Analysis and Machine Intelligence, vol. 48, no. 4, pp. 4605–4622, 2026.DOI
[65]S. Li, Q. Xu, Z. Yang, Z. Wang, L. Zhang, X. Cao, and Q. Huang, “Focal-SAM: Focal sharpness-aware mini- mization for long-tailed classification,” in International Conference on Machine Learning, 2025, pp. 36 624–36 651.arxiv.org/abs/2505.01660
[66]L. van der Maaten and G. Hinton, “Visualizing data using t-SNE,” J. Mach. Learn. Res., vol. 9, pp. 2579–2605, 2008.