Pith. sign in

REVIEW 3 major objections 6 minor 55 references

Prior-Constrained Association Learning for Fine-Grained Generalized Category Discovery

T0 review · 3 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A hard prior constraint on grouping lifts fine-grained GCD accuracy to a new state of the art.

desk verdict A genuinely new association constraint with strong ablations, but the reported margins are likely inflated by per-dataset test-set tuning. read the letter →

arxiv 2502.09501 v1 pith:R2WAL6JE submitted 2025-02-13 cs.CV

Menglin Wang , Zhun Zhong , Xiaojin Gong This is my paper · ORCID
classification cs.CV
keywords generalizedcategorydiscoveryfine-grainedvisualrecognitionassociationlearningprototypicalcontrastivesemi-supervisedclusteringnovelnon-parametricclassification
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper proposes a method for Generalized Category Discovery (GCD), where unlabeled images may belong to known or novel classes and a labeled subset exists for each known class. Its central claim is that the labeled prior should be enforced as a hard constraint during the data-association step, not just as a pre- or post-clustering refinement, and that this yields more reliable pseudo-groups. Those groups become prototypes in non-parametric prototypical contrastive learning, which can be combined with a parametric classifier to mutually boost both. On fine-grained benchmarks, the method reports state-of-the-art accuracy, for example 72.6% on CUB and 72.2% on Stanford Cars, improving over the previous best by 4.4 and 15.3 percentage points respectively. A reader should care because fine-grained GCD is closer to real-world open-set recognition, where classes differ subtly and novel categories matter.

What carries the argument

The load-bearing component is the prior-constrained greedy association algorithm (Algorithm 1), which takes as input a distance matrix, a threshold epsilon, and the number of known categories. It initializes one group per known class, masks pairwise distances among known-class proxies, and iterates over candidate pairs in ascending distance order, merging two groups only if doing so would not place two different known classes in the same group. This per-step validity check is what distinguishes the method from pre/post-refined clustering, and it is what produces pseudo-groups that respect the labeled ground truth. These groups serve as prototypes in a non-parametric prototypical contrastive loss, and optionally as an auxiliary loss alongside a parametric classifier in two-stage training.

What would settle it

Train the model on a fine-grained dataset while varying epsilon over a grid, choosing the best value on the test set as done in the paper, and cross-check by selecting epsilon on a held-out validation split of known classes; if the validation-selected threshold yields substantially lower accuracy, the reported numbers depend on test-set tuning rather than a principled threshold choice.

Watch

Extended reading notes

Core claim

The paper's central discovery is that enforcing the constraint that no discovered group contains more than one known class, at every merge step of a greedy association process, prevents both direct and indirect false associations that earlier pre/post refinement schemes tolerate. Starting from per-class proxies for labeled data and unlabeled instance features, the algorithm computes pairwise Jaccard distances, sorts candidate pairs below a threshold, and merges groups in ascending distance order while checking the prior constraint at each step. The resulting instance groups are used as prototypes in a non-parametric prototypical contrastive loss, pulling each instance toward its group centroid and pushing it away from other centroids. The paper further shows that this non-parametric classification can be combined with a parametric classifier through a two-stage training procedure, where the non-parametric loss warms up the backbone first and then both losses are optimized together. The joint model achieves the reported gains, with particularly large improvements on novel classes.

Load-bearing premise

The method assumes that the pairwise Jaccard distances in the current feature space reliably reflect semantic similarity, and that a single global threshold epsilon correctly separates true from false pairs across both known and novel classes, with the threshold and subset association ratio set per dataset without a described validation procedure.

Editorial extensions

If this is right

  • If the claimed gains hold, GCD methods should treat the labeled prior as a hard constraint during association rather than as a post-hoc refinement, and this principle should transfer to other clustering-with-partial-labels tasks.
  • The non-parametric association-based framework alone, without knowing the ground-truth class number, reaches competitive accuracy, suggesting that class-number-free settings can benefit from the same design.
  • Combining non-parametric and parametric classifiers with a warm-up stage is a viable recipe for further gains when the class number is known.
  • Subset association provides a practical route to scaling the greedy association to datasets with many images per category, at some cost in accuracy when per-category images are few.
  • The consistent improvement on 'New' classes indicates the constraint does not sacrifice novel-category discovery while protecting old-class purity.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The hard constraint could be adapted to other partial-label settings, such as when labeled data comes from a subset of classes but not necessarily the same feature space, although the paper does not explore this.
  • Because the epsilon threshold and subset ratio are dataset-specific and no validation procedure is reported, the method may be sensitive to these hyperparameters in deployment; a principled selection scheme would strengthen the claim.
  • The largest gains appear on fine-grained datasets where DINO features are initially weak (Stanford Cars rises from 12.9 to 70.5 before and after training), suggesting the association constraint is particularly helpful when the feature space starts poor; this could motivate combining the constraint with weaker backbones.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 6 minor

Summary. The paper proposes Prior-Constrained Association Learning (PAL) for fine-grained generalized category discovery. The method computes pairwise Jaccard distances between labeled class proxies and unlabeled instances, then greedily merges pairs below a threshold epsilon while preventing any group from containing more than one known-class proxy. The resulting groups serve as prototypes in a non-parametric prototypical contrastive loss, and can optionally be combined with a SimGCD-style parametric classifier in a two-stage procedure. Experiments are conducted on four fine-grained datasets and two generic datasets. The paper reports state-of-the-art accuracy on CUB (72.6%, non-parametric evaluation), Stanford Cars (72.2%), and Aircraft (58.8%), with smaller gains on Herbarium19 and generic datasets, plus a class-number-unknown setting.

Significance. If the reported results are reliable, the paper makes a meaningful empirical contribution to fine-grained GCD: the association-based non-parametric classifier alone outperforms prior methods on several benchmarks, and the joint parametric/non-parametric training yields consistent improvements. The work is also credibly positioned relative to DCCL, PromptCAL, SimGCD, and CMS, and it ships an ablation study covering association algorithms, subset size, threshold, loss weight, and reduced-label settings. The central weakness is that the headline gains depend on dataset-specific association hyperparameters that appear to be selected using test-set accuracy, which is especially concerning for the Stanford Cars result. This issue must be resolved before the SOTA claim can be accepted.

major comments (3)
  1. [Appendix A.2 and Figure 4] The association threshold epsilon and the subset-association ratio are set per dataset without any described validation split: epsilon is 0.35 for CUB/Cars/Aircraft and 0.6 for Herbarium19/Cifar100/ImageNet-100, and subset ratios are 50%/30%/30% for Aircraft/Cifar100/ImageNet-100. Figure 4 sweeps epsilon and plots test accuracy, which is consistent with selecting hyperparameters on the test benchmark itself. This is load-bearing because the association algorithm produces both the training pseudo-labels and the final evaluation labels, and because the gains are largest exactly where the initial feature space is weakest (Stanford Cars goes from 12.9% before training to 70.5% after, Table 8). Please provide a validation-based selection procedure, report results under fixed hyperparameters across datasets, or demonstrate that the reported margins are robust to threshold perturbations within a plausible range.
  2. [Section 3.3 and Algorithm 1] The claim that the generated groups are 'guaranteed to adhere to the ground truth of labeled subset' is stronger than what the algorithm actually enforces. The constraint only prevents merging two groups that already contain different known-class proxies; in Algorithm 1, line 22-25 permits merging any two groups when at least one group is unlabeled (grpLabel >= C1), so an unlabeled instance from a known class can still be absorbed into a group that contains a proxy of a different known class. Thus the prior does not fully preserve the known-class structure, and the self-training loop can reinforce such errors without bound. This does not invalidate the method, but the stated guarantee should be softened or, preferably, supported with an analysis of the conditions under which the constraint bounds known-class contamination.
  3. [Table 2 and Table 9] The main SOTA claims in Table 2 are reported as single numbers, while error bars are provided only for the non-parametric-only, class-number-unknown setting in Table 9. Given that the proposed method uses stochastic components (random subset association, mini-batch sampling) and that the headline margins over CMS are 4.4% and 15.3%, please report mean and standard deviation over at least three runs for the full model on the fine-grained datasets, and state whether the same association hyperparameters are used in every run.
minor comments (6)
  1. [Section 3.4] Typographical error: 'effectivestand-alone' should be 'effective stand-alone'.
  2. [Appendix A.2] Typographical error: 'for For both two training stages' should be 'For both two training stages'.
  3. [Table 3] The abbreviation 'A&A' in the header is not defined in the caption; the caption later mentions 'Association and Assign', but the table should spell out the abbreviation for readability.
  4. [Table 4] The row label 'Semi-DBSCANw/ constraint' is missing a space; it should be 'Semi-DBSCAN w/ constraint'.
  5. [Equation (1)] The summation bounds are typeset as 'PC−1 j=0' in the text; please format the sum with proper lower and upper limits for clarity.
  6. [Figure 4] The threshold analysis plot would benefit from error bars and from a marker or vertical line indicating the chosen epsilon for each dataset, so that the reader can see the operating point relative to the peak.

Circularity Check

1 steps flagged · score 6.0 of 10

Headline SOTA gains depend on association hyperparameters (epsilon, subset ratio, beta) selected on the test benchmarks without a validation split; the derivation itself is otherwise self-contained.

  1. fitted input called prediction [Appendix A.2 (More Implementation Details); Appendix A.4, Figure 4; Table 7]
    "The association threshold ϵ is set as 0.35 for all fine-grained datasets except Herbarium19, on which we use a larger threshold of 0.6 as generic datasets Cifar100 and ImageNet-100. ... Due to the number of larger per-category images, subset association is performed on Aircraft, Cifar100 and ImageNet-100 with subset size as 50%, 30% and 30% respectively. ... Figure 4 plots the model accuracy under varying association thresholds. We observe that accuracy in 'New' and 'All' classes share a consistent trend, and their peak accuracy appears at a similar threshold."

    The per-dataset values of the association threshold epsilon, the subset-association ratio, and the second-stage loss weight beta are selected by inspecting the same test-benchmark ACC curves (Figure 4 sweeps epsilon and plots ACC; Table 7 sweeps beta on Stanford Cars) with no validation split described. The headline SOTA numbers in Table 2 (e.g., 72.6 CUB, 72.2 Stanford Cars) are then reported at the selected hyperparameter values. This is a fitted-input-called-prediction loop: the test labels are used to choose the association hyperparameters, and the resulting test accuracy is presented as the method's performance. The reported gains are therefore partly constructed by test-set selection rather than derived from the association algorithm alone.

full rationale

The core methodological derivation is not definitionally circular. The prior-constrained greedy association is a semi-supervised cannot-link agglomerative procedure: labeled old-class proxies are kept in distinct groups and the merge rule forbids merging two old-proxy groups, which is a legitimate use of labeled prior. The non-parametric prototypical contrastive loss (Eq. 1) trains features to be close to the centroids of the association's groups; the association is recomputed each epoch from the current features. This is an iterative self-training loop, but it is not a formal reduction of the claimed result to its inputs: the final evaluation compares the association's pseudo-labels to external ground-truth labels after Hungarian matching, and Table 8 shows that a poor initial association (12.9% on Stanford Cars) can be substantially improved, which would not happen if the loop merely reproduced its own input. The parametric branch is an external baseline (SimGCD) and is not self-cited. The one severe circularity-adjacent flaw is the experimental protocol: the association threshold epsilon, subset-association ratios, and the loss weight beta are dataset-specific and selected by inspecting test-benchmark accuracy (Figure 4 and Table 7), with no validation split described in Appendix A.2. Because these hyperparameters directly control the greedy merging and the final association-based assignment, the headline SOTA numbers are partly fitted to the test set. This is a fitted-input-called-prediction pattern and raises the circularity score to 6, even though the method has independent algorithmic content. The self-citations in the paper (Wang et al. 2021a,b; Zhong et al. 2017) are technical references for re-ranking/Jaccard distance and prototypical contrast, not load-bearing uniqueness claims.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central claim rests on the reliability of feature-space distances, a global threshold, and a self-training loop. No new physical or conceptual entities are introduced. The main free parameters are dataset-specific (threshold, subset ratio, loss weight), and the paper does not describe a validation-based selection protocol.

free parameters (3)
  • association threshold epsilon = 0.35 (CUB, Stanford Cars, Aircraft); 0.6 (Herbarium19, Cifar100, ImageNet-100)
    Controls which candidate pairs are considered for greedy merging; Figure 4 shows accuracy varies with epsilon, and the value is set per dataset.
  • subset association ratio = 100% (CUB, Stanford Cars, Herbarium19); 50% Aircraft; 30% Cifar100 and ImageNet-100
    The association is run on a random subset of unlabeled images for datasets with many images per category; Table 5 shows significant accuracy differences.
  • loss weight beta for non-parametric loss in stage 2 = 0.1
    Weights the prototypical contrastive loss when combined with SimGCD; Table 7 shows sensitivity.
assumptions (4)
  • domain assumption DINO-pretrained ViT-B/16 provides a feature space in which pairwise Jaccard distances reflect semantic similarity for fine-grained categories.
    The association and contrastive learning take place in this feature space (Section 3.3, Implementation Details).
  • ad hoc to paper A global distance threshold can separate same-category pairs from different-category pairs well enough for greedy merging to produce useful pseudo-labels.
    The algorithm's candidate selection uses a single epsilon; no per-class or local threshold is used (Algorithm 1, Figure 4).
  • domain assumption Self-training with model-generated pseudo-labels improves representation and does not collapse, even when initial association accuracy is low (e.g., 12.9% on Stanford Cars before training).
    The method iteratively re-runs association and trains with prototypical contrastive loss (Section 3.4, Table 8).
  • ad hoc to paper Enforcing that each group contains at most one known class is sufficient to preserve the known-class structure during association.
    The prior constraint only blocks merges that would put two known-class proxies in one group; it does not directly prevent unlabeled instances from being assigned to the wrong known class or novel classes from merging (Section 3.3).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Prior-Constrained Association Learning for Fine-Grained Generalized Category Discovery." pith.science (2026). https://pith.science/paper/R2WAL6JE

@misc{pith2026250209501,
  author       = {Pith},
  title        = {Pith review of: Prior-Constrained Association Learning for Fine-Grained Generalized Category Discovery},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/R2WAL6JE}},
  note         = {Machine review of arXiv:2502.09501}
}
read the original abstract

This paper addresses generalized category discovery (GCD), the task of clustering unlabeled data from potentially known or unknown categories with the help of labeled instances from each known category. Compared to traditional semi-supervised learning, GCD is more challenging because unlabeled data could be from novel categories not appearing in labeled data. Current state-of-the-art methods typically learn a parametric classifier assisted by self-distillation. While being effective, these methods do not make use of cross-instance similarity to discover class-specific semantics which are essential for representation learning and category discovery. In this paper, we revisit the association-based paradigm and propose a Prior-constrained Association Learning method to capture and learn the semantic relations within data. In particular, the labeled data from known categories provides a unique prior for the association of unlabeled data. Unlike previous methods that only adopts the prior as a pre or post-clustering refinement, we fully incorporate the prior into the association process, and let it constrain the association towards a reliable grouping outcome. The estimated semantic groups are utilized through non-parametric prototypical contrast to enhance the representation learning. A further combination of both parametric and non-parametric classification complements each other and leads to a model that outperforms existing methods by a significant margin. On multiple GCD benchmarks, we perform extensive experiments and validate the effectiveness of our proposed method.

Figures

Figures reproduced from arXiv: 2502.09501 by the authors.

Figure 1
Figure 1. An illustration of implicit false association, and how our proposed association avoids it. Arrowed line represents [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. An overview of our method. Non-parametric classification is in the form of prototypical contrastive learning, with [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Estimated class number on each dataset. When is subset association necessary? For large-scale datasets or when number of images per category is high, we propose to perform the subset association. To verify the ef￾fect of subset association, we provide in [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Analysis on association threshold ϵ. Accuracy evaluated by association before and after training. To gain a more clear idea of the performance boost, [PITH_FULL_IMAGE:figures/full_fig_p011_4.png]
Figure 5
Figure 5. Figure 5: Visualization of features extracted by different models on CUB dataset. [PITH_FULL_IMAGE:figures/full_fig_p012_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

55 extracted references · 47 canonical work pages

  1. [1]

    , " * write output.state after.block = add.period write newline

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts #0 'before.a...

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    Bair, E. 2013. Semi-supervised clustering methods. Wiley Interdisciplinary Reviews: Computational Statistics

    work page 2013

  4. [4]

    Basu, S. 2002. Semi-supervised clustering by seeding. In Internation Conference on Machine Learning

    work page 2002

  5. [5]

    Bilenko, M.; Basu, S.; and Mooney, R. J. 2004. Integrating constraints and metric learning in semi-supervised clustering. In Proceedings of the twenty-first international conference on Machine learning

    work page 2004

  6. [6]

    Caron, M.; Touvron, H.; Misra, I.; J \'e gou, H.; Mairal, J.; Bojanowski, P.; and Joulin, A. 2021. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision

    work page 2021

  7. [7]

    Chen, T.; Kornblith, S.; Norouzi, M.; and Hinton, G. 2020. A Simple Framework for Contrastive Learning of Visual Representations. In International Conference on Machine Learning

    work page 2020

  8. [8]

    I.; Mitiche, A.; and Ben Ayed, I

    Chiaroni, F.; Dolz, J.; Masud, Z. I.; Mitiche, A.; and Ben Ayed, I. 2023. Parametric information maximization for generalized category discovery. In Proceedings of the IEEE/CVF International Conference on Computer Vision

    work page 2023

Show all 55 references
  1. [9]

    Choi, S.; Kang, D.; and Cho, M. 2024. Contrastive Mean-Shift Learning for Generalized Category Discovery. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    2024

  2. [10]

    Davidson, I.; and Ravi, S. 2005. Clustering with constraints: Feasibility issues and the k-means algorithm. In Proceedings of the SIAM international conference on data mining. SIAM

    2005

  3. [11]

    Deng, J.; Dong, W.; Socher, R.; Li, L.-J.; Li, K.; and Fei-Fei, L. 2009. ImageNet: A Large-Scale Hierarchical Image Database. In CVPR

    2009

  4. [12]

    Dosovitskiy, A.; Beyer, L.; Kolesnikov, A.; Weissenborn, D.; Zhai, X.; Unterthiner, T.; Dehghani, M.; Minderer, M.; Heigold, G.; Gelly, S.; et al. 2021. An image is worth 16x16 words: Transformers for image recognition at scale. International Conference on Learning Representations

    2021

  5. [13]

    Ester, M.; Kriegel, H.-P.; Sander, J.; Xu, X.; et al. 1996. A density-based algorithm for discovering clusters in large spatial databases with noise. In Kdd

    1996

  6. [14]

    Fei, Y.; Zhao, Z.; Yang, S.; and Zhao, B. 2022. Xcon: Learning with experts for fine-grained category discovery. British Machine Vision Conference

    2022

  7. [15]

    Fini, E.; Sangineto, E.; Lathuiliere, S.; Zhong, Z.; Nabi, M.; and Ricci, E. 2021. A unified objective for novel class discovery. In Proceedings of the IEEE/CVF International Conference on Computer Vision

    2021

  8. [16]

    Ge, Y.; Chen, D.; Zhu, F.; Zhao, R.; and Li, H. 2020. Self-paced Contrastive Learning with Hybrid Memory for Domain Adaptive Object Re-ID. In Advances in Neural Information Processing Systems

    2020

  9. [17]

    Gutmann, M.; and Hyvärinen, A. 2010. Noise-contrastive estimation: A new estimation principle for unnormalized statistical models. In AISTATS

    2010

  10. [18]

    Han, K.; Rebuffi, S.-A.; Ehrhardt, S.; Vedaldi, A.; and Zisserman, A. 2021. Autonovel: Automatically discovering and learning novel visual categories. IEEE Transactions on Pattern Analysis and Machine Intelligence

    2021

  11. [19]

    A.; and Wong, M

    Hartigan, J. A.; and Wong, M. A. 1979. Algorithm AS 136: A k-means clustering algorithm. Journal of the royal statistical society. series c (applied statistics)

    1979

  12. [20]

    He, K.; Fan, H.; Wu, Y.; Xie, S.; and Girshick, R. 2020. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition

    2020

  13. [21]

    Johnson, S. C. 1967. Hierarchical clustering schemes. Psychometrika

    1967

  14. [22]

    Khosla, P.; Teterwak, P.; Wang, C.; Sarna, A.; Tian, Y.; Isola, P.; Maschinot, A.; Liu, C.; and Krishnan, D. 2020. Supervised Contrastive Learning. arXiv preprint arXiv:2004.11362

    2020 arXiv

  15. [23]

    Kim, H.; Suh, S.; Kim, D.; Jeong, D.; Cho, H.; and Kim, J. 2023. Proxy anchor-based unsupervised learning for continuous generalized category discovery. In Proceedings of the IEEE/CVF International Conference on Computer Vision

    2023

  16. [24]

    D.; and Manning, C

    Klein, D.; Kamvar, S. D.; and Manning, C. D. 2002. From instance-level constraints to space-level constraints: Making the most of prior knowledge in data clustering. In International Conference on Machine Learning

    2002

  17. [25]

    Krause, J.; Stark, M.; Deng, J.; and Fei-Fei, L. 2013. 3d object representations for fine-grained categorization. In Proceedings of the IEEE international conference on computer vision workshops

    2013

  18. [26]

    Krizhevsky, A.; Hinton, G.; et al. 2009. Learning multiple layers of features from tiny images

    2009

  19. [27]

    Kuhn, R. W. 1955. The hungarian method for the assignment problem. Naval research logistics quarterly

    1955

  20. [28]

    H.; Jain, A

    Lange, T.; Law, M. H.; Jain, A. K.; and Buhmann, J. M. 2005. Learning with constrained and unlabelled data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. IEEE

    2005

  21. [29]

    Li, J.; Zhou, P.; Xiong, C.; and Hoi, S. 2021. Prototypical Contrastive Learning of Unsupervised Representations. In International Conference on Learning Representations

    2021

  22. [30]

    MacQueen, J.; et al. 1967. Some methods for classification and analysis of multivariate observations. In Proceedings of the fifth Berkeley symposium on mathematical statistics and probability. Oakland, CA, USA

    1967

  23. [31]

    Maji, S.; Rahtu, E.; Kannala, J.; Blaschko, M.; and Vedaldi, A. 2013. Fine-grained visual classification of aircraft. arXiv preprint arXiv:1306.5151

    2013 arXiv

  24. [32]

    Murtagh, F.; and Contreras, P. 2012. Algorithms for hierarchical clustering: an overview. Wiley Interdisciplinary Reviews: Data Mining and Knowledge Discovery

    2012

  25. [33]

    A.; Cubuk, E

    Oliver, A.; Odena, A.; Raffel, C. A.; Cubuk, E. D.; and Goodfellow, I. 2018. Realistic evaluation of deep semi-supervised learning algorithms. Advances in neural information processing systems

    2018

  26. [34]

    Pu, N.; Zhong, Z.; and Sebe, N. 2023. Dynamic Conceptional Contrastive Learning for Generalized Category Discovery. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    2023

  27. [35]

    Rosvall, M.; and Bergstrom, C. T. 2008. Maps of random walks on complex networks reveal community structure. Proceedings of the national academy of sciences

    2008

  28. [36]

    Ruiz, C.; Spiliopoulou, M.; and Menasalvas, E. 2010. Density-based semi-supervised clustering. Data mining and knowledge discovery

    2010

  29. [37]

    Sun, Y.; and Li, Y. 2022. Opencon: Open-world contrastive learning. Transactions on Machine Learning Research

    2022

  30. [38]

    C.; Liu, Y.; Ambrose, B.; Tulig, M.; and Belongie, S

    Tan, K. C.; Liu, Y.; Ambrose, B.; Tulig, M.; and Belongie, S. 2019. The herbarium challenge 2019 dataset. Workshop on Fine-Grained Visual Categorization

    2019

  31. [39]

    Vaze, S.; Han, K.; Vedaldi, A.; and Zisserman, A. 2021. Open-set recognition: A good closed-set classifier is all you need? arXiv preprint arXiv:2110.06207

    2021 arXiv

  32. [40]

    Vaze, S.; Han, K.; Vedaldi, A.; and Zisserman, A. 2022. Generalized category discovery. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    2022

  33. [41]

    Vaze, S.; Vedaldi, A.; and Zisserman, A. 2024. No Representation Rules Them All in Category Discovery. Advances in Neural Information Processing Systems

    2024

  34. [42]

    Wagstaff, K.; Cardie, C.; Rogers, S.; Schr \"o dl, S.; et al. 2001. Constrained k-means clustering with background knowledge. In Internation Conference on Machine Learning

    2001

  35. [43]

    Wah, C.; Branson, S.; Welinder, P.; Perona, P.; and Belongie, S. 2011. The caltech-ucsd birds-200-2011 dataset

    2011

  36. [44]

    Wang, G.; Zhan, Y.; Wang, X.; Song, M.; and Nahrstedt, K. 2022. Hierarchical semi-supervised contrastive learning for contamination-resistant anomaly detection. In European Conference on Computer Vision. Springer

    2022

  37. [45]

    Wang, H.; Vaze, S.; and Han, K. 2024. SPTNet: An Efficient Alternative Framework for Generalized Category Discovery with Spatial Prompt Tuning. In The Twelfth International Conference on Learning Representations

    2024

  38. [46]

    Wang, M.; Lai, B.; Chen, H.; Huang, J.; Gong, X.; and Hua, X.-S. 2021 a . Towards Precise Intra-camera Supervised Person Re-Identification. In IEEE/CVF Winter Conference on Applications of Computer Vision

    2021

  39. [47]

    Wang, M.; Lai, B.; Huang, J.; Gong, X.; and Hua, X.-S. 2021 b . Camera-aware Proxies for Unsupervised Person Re-Identification. In AAAI Conference on Artificial Intelligence

    2021

  40. [48]

    Wen, X.; Zhao, B.; and Qi, X. 2023. Parametric classification for generalized category discovery: A baseline study. In Proceedings of the IEEE/CVF International Conference on Computer Vision

    2023

  41. [49]

    X.; and Lin, D

    Wu, Z.; Xiong, Y.; Yu, S. X.; and Lin, D. 2018. Unsupervised Feature Learning via Non-Parametric Instance Discrimination. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    2018

  42. [50]

    Xiao, T.; Li, S.; Wang, B.; Lin, L.; and Wang, X. 2017. Joint Detection and Identification Feature Learning for Person Search. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    2017

  43. [51]

    J.; and Ng, A

    Xing, E.; Jordan, M.; Russell, S. J.; and Ng, A. 2002. Distance metric learning with application to clustering with side-information. Advances in neural information processing systems

    2002

  44. [52]

    Yin, X.; Chen, S.; Hu, E.; and Zhang, D. 2010. Semi-supervised clustering with metric learning: An adaptive kernel method. Pattern Recognition

    2010

  45. [53]

    Zhang, S.; Khan, S.; Shen, Z.; Naseer, M.; Chen, G.; and Khan, F. S. 2023. Promptcal: Contrastive affinity learning via auxiliary prompts for generalized novel category discovery. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    2023

  46. [54]

    Zhao, B.; Wen, X.; and Han, K. 2023. Learning Semi-supervised Gaussian Mixture Models for Generalized Category Discovery. Proceedings of the IEEE/CVF International Conference on Computer Vision

    2023

  47. [55]

    Zhong, Z.; Zheng, L.; and Li, S. 2017. Re-ranking Person Re-identification with k-Reciprocal Encoding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    2017

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.