Pith. sign in

REVIEW 3 major objections 6 minor 64 references

A Dirichlet-process evidence comparison decides whether each streamed sample joins a known category or starts a new one, and the paper reports best overall accuracy on 8 of 10 benchmarks among methods without external knowledge.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-02 04:59 UTC pith:THAO2OKZ

load-bearing objection DP-BOA is a genuinely new, well-engineered answer to the OCD birth-or-assign problem, and its gains are credible but rest on an assumption—novel classes share the known classes' Gaussian geometry—that the authors honestly show breaks down under domain shift. the 3 major comments →

arxiv 2607.13504 v1 pith:THAO2OKZ submitted 2026-07-15 cs.CV

DP-BOA: Dirichlet-Process Birth-or-Assign for On-the-Fly Category Discovery

Peiyan Gu , Zixin Teng , Xuming He This is my paper
classification cs.CV
keywords on-the-fly category discoveryDirichlet processGaussian mixture modelNormal-Inverse-Wishart priorposterior predictivestreaming classificationopen-world recognitionnovel class discovery
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper claims that the hardest choice in on-the-fly category discovery — assign this test sample to an existing category or declare a new one — should be made as a Bayesian evidence comparison rather than as a fallback from a failed match. It builds an online Dirichlet-process Gaussian mixture model whose 'birth' hypothesis has its own posterior-predictive density, calibrated from labeled known classes, and compares that against the predictive density of every existing category weighted by its current size. Because category posteriors are updated online with closed-form Normal-Inverse-Wishart updates, predictive uncertainty contracts as evidence accumulates and category geometry is captured by full covariance matrices. The paper reports that this consistently outperforms strong baselines on standard OCD benchmarks, with particularly large gains on novel-class discovery, and that each component of the design is supported by ablations.

Core claim

The core claim is that decision rule (12) — choose the maximum of α·t_d(z_t | θ0) for birth versus max_k n_k·t_d(z_t | θ_k) for assignment — turns category birth into an explicit, evidence-bearing alternative. Here t_d is a multivariate Student-t posterior predictive obtained by integrating out Gaussian parameters under a Normal-Inverse-Wishart prior, n_k counts samples already in category k, and α is the Dirichlet-process concentration. The paper argues this replaces Hamming-ball or radius thresholds with a statistically grounded comparison: the birth hypothesis is the prior predictive of a shared NIW prior estimated from labeled known classes, while assignment hypotheses are size-weighted

What carries the argument

The central object is an online Dirichlet-process Gaussian mixture with a Normal-Inverse-Wishart prior. The DP supplies the discrete prior over category indices (existing category k has probability n_k/(α+N), a new category has α/(α+N)); the NIW prior supplies closed-form conjugate updates that turn per-category sufficient statistics (count, mean, scatter) into multivariate Student-t predictive densities. An empirical-Bayes step calibrates the shared prior from labeled known-class statistics (global mean, pooled within-class covariance, covariance of class means), and a data-scaled heuristic sets the degrees of freedom. The decision compares prior-weighted predictives, so birth is no longer

Load-bearing premise

The load-bearing assumption is that novel-class features in the stream look like Gaussian-shaped clouds whose covariance scale and orientation are reasonably predicted by the shared prior calibrated from labeled known classes; if a genuinely new category has a very different shape or domain, the birth evidence is miscalibrated and known-class accuracy drops (the paper's own domain-shift test shows this).

What would settle it

A concrete test: on a labeled stream where novel classes are constructed to be multi-modal or domain-shifted (e.g., known classes from one domain, novel classes from a visually different domain), compute the empirical covariance of each novel class and compare it with the prior scale Ψ0 set from the pooled within-class covariance. If novel-class scales differ by more than a factor of two, DP-BOA should produce false births or missed births; the paper's DomainNet Clipart result (known accuracy 23.7) already exhibits the predicted collapse.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Streaming systems that must spot new categories (robotics, content moderation) get a decision rule whose confidence calibrates automatically as each category accumulates evidence.
  • Novel-class discovery improves without external knowledge: on CUB the paper reports 51.6% novel accuracy versus 40.9% for the strongest compared method that uses language assistance.
  • Known-class accuracy stays competitive, so the method does not sacrifice stability for discovery.
  • A low-rank variant, DP-BOA-L, keeps most of the accuracy while cutting head memory from roughly 0.6 GB to 50 MB, making full-covariance Bayesian scoring practical at larger scale.
  • The decision rule is modular: the Gaussian likelihood could be swapped for richer category models without changing the birth-or-assign comparison.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same posterior-predictive comparison could serve as a streaming out-of-distribution or novelty detector, since a high birth score is essentially evidence that no existing category explains the sample.
  • The paper's domain-shift stress test suggests the single-Gaussian birth density is the fragile link; a testable extension would replace the shared NIW prior with a mixture of priors or adapt the prior online from recently birthed categories.
  • The data-scaled n0 heuristic (half the average known-class size, capped) hints that scale calibration, not absolute thresholds, carries most of the method's robustness; one could test this by varying known-class support size and checking whether the optimal n0 tracks the heuristic.
  • Because decisions are made from sufficient statistics alone, the method could be deployed in a fully distributed or federated stream where each node maintains only counts, means, and scatters.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The paper proposes DP-BOA, a streaming category-discovery method that frames each on-the-fly decision as a Bayesian comparison between assigning a sample to an existing category and birthing a new category. The model is an online Dirichlet-process Gaussian mixture with a Normal-Inverse-Wishart prior; labeled support categories are used to warm-start known-category posteriors and to calibrate a shared NIW prior via empirical-Bayes moment matching. At test time, each feature is scored by predictive Student-t densities for existing categories and for a new category, weighted by DP predictive probabilities (Eq. (12)), and the chosen category's sufficient statistics are updated online. Experiments on ten OCD benchmarks report state-of-the-art or competitive accuracy, especially on novel classes, with ablations, sensitivity analyses, runtime/memory measurements, and a low-rank variant DP-BOA-L.

Significance. If the reported results hold under a fully valid evaluation protocol, DP-BOA is a meaningful step forward for on-the-fly category discovery. Its strengths are that the probabilistic derivation is standard and carefully documented (NIW updates in Appendix C, Student-t predictives in Appendix D, DP predictive rule in Appendix E); the method does not use external knowledge; the empirical study is broad, with ablations, standard deviations, cluster-growth diagnostics, temporal diagnostics, and robustness checks; and the same-feature comparison with simpler probabilistic heads in Appendix N helps isolate the contribution of the explicit birth hypothesis. The central idea — treating birth as a posterior-predictive alternative rather than a rejection fallback — is conceptually clean and is supported by the ablations. The main open risks are the selection of the two global hyperparameters (α and n_cap) and the load-bearing assumption that novel-class geometry matches the known-class-calibrated NIW prior.

major comments (3)
  1. [Sec. 4.1 and Sec. 4.3, Tables 4–5] The global defaults α=1e-9 and n_cap=50 appear to be selected after inspecting sensitivity tables on the same test benchmarks. Since Eq. (12) contains α as a multiplicative birth prior and n_cap controls n0, hence the birth predictive scale and tail behavior, this is tuning on the test set. Table 4 shows that n0=25 changes OxfordPets All accuracy by +1.8 points relative to the default, and Table 15 shows n_cap=25 vs 50 changes CIFAR-100 Novel from 26.0 to 32.0. The manuscript should either tune these values on a held-out validation split (as Appendix N does for the simpler heads) or report the full grid and a pre-specified selection rule that does not use benchmark labels. This is load-bearing for the claim of consistent state-of-the-art performance.
  2. [Sec. 3.4, Eq. (12), and Appendix M.2–M.3] The birth evidence is the NIW prior predictive calibrated exclusively from labeled known classes (Eqs. 7–11), and each category is a single full-covariance Gaussian. Novel classes are therefore assumed to live in the same Gaussian/NIW geometry as known classes. Appendix M.3 itself shows that when this assumption is violated by domain shift, the mechanism degrades: target Clipart Known accuracy is 23.7 vs 31.6 for PHE, and target Clipart Novel is 14.1. On the standard benchmarks, Appendix M.2 offers only qualitative 2D PCA contour plots, not a quantitative check that novel-class covariance scale and anisotropy match the prior. Because Eq. (12) is a ratio between birth and existing-category densities, miscalibration of the birth prior directly changes the birth rate and hence the reported novel-class accuracy. Please add a quantitative diagnostic on the standard benchmarks (e.g., comparing
  3. [Sec. 4.2, Tables 1–2] The abstract and Sec. 4.2 claim DP-BOA 'consistently outperforms strong baselines,' but on ImageNet100, DP-BOA's All accuracy (33.8) is slightly below PHE (34.0), and on OxfordPets, Sync achieves higher All (61.6 vs 59.0) and Novel (57.5 vs 56.6). While Sync uses external knowledge and the comparison is conservative, the claim should be qualified to 'among purely visual OCD methods' and should not imply universal superiority over all compared methods. This is a presentation issue, but it affects how the results are interpreted.
minor comments (6)
  1. [Eq. (12) and Sec. 3.5] Eq. (12) omits the common factor 1/(α+N_{t−1}) and the max over existing categories; this is fine for argmax, but the notation should be clarified so readers do not think the birth term is being compared on a different scale.
  2. [Abstract and Sec. 1] The project page is referred to only as 'DP-BOA'; a URL should be provided.
  3. [Sec. 4.1 and Appendix B] Standard deviations are reported only in the appendix. Given that some SOTA gaps are small (e.g., ImageNet100 All: 33.8 vs 34.0), main-table error bars or significance statements would help.
  4. [Appendix M.2] The statement that 'many categories exhibit a dominant central region' is based on visual inspection. A quantitative summary (e.g., fraction of categories passing a multivariate normality test in PCA space, or coverage of 1σ/2σ contours) would make the claim reproducible.
  5. [Sec. 3.4, Eq. (11)] The definition of n̄ as an average over known classes should be restated near Eq. (11) for readability; currently it is only clear from the surrounding text.
  6. [Appendix N] The held-out validation split used to tune thresholds for the simpler heads is not described in enough detail. Please state how it is carved from the training data and whether the same split was used for any DP-BOA hyperparameter selection.

Circularity Check

0 steps flagged

No derivation-level circularity; the only self-citations are non-load-bearing related-work mentions, and the birth prior is calibrated on labeled external supervision.

full rationale

DP-BOA's decision rule (Eq. 12) assembles standard DP predictive probabilities (Eqs. 5–6), closed-form Student-t predictives under an NIW prior (Eqs. 3–4; App. D), and an empirical-Bayes calibration of (μ0, Ψ0, κ0) from the labeled support set (Sec. 3.4, Eqs. 7–11; App. F). None of these inputs is defined in terms of the novel-class labels or the reported accuracies; the novel-class hypothesis uses the prior predictive of the same NIW hierarchy, with scale set by known-class statistics plus fixed hyperparameters α=1e-9 and ncap=50. The paper's own domain-shift stress test (App. M.3, Table 16) shows that the known-class geometry assumption is load-bearing and degrades under target shift (Clipart Known 23.7), but that is an assumption/robustness limitation, not a circular reduction. The only self-citations, [56,57] in a related-work list of clustering objectives, are not load-bearing for the central claim. Hyperparameter defaults were selected with sensitivity analyses on benchmarks (Tabs. 4–5, 15); this is a tuning concern, not evidence that Eq. (12) reduces to its inputs. Overall, the derivation is self-contained: no fitted parameter is renamed as a prediction, and no author-specific uniqueness theorem or ansatz is imported.

Axiom & Free-Parameter Ledger

2 free parameters · 7 axioms · 0 invented entities

The central claim rests on standard Bayesian machinery (DP, NIW, Student-t marginals) plus two domain assumptions: novel categories have the same Gaussian/NIW geometry as known classes, and a single full-covariance Gaussian adequately models category features. Two free hyperparameters (α, ncap) are fit to benchmark outcomes. No fundamentally new physical or metaphysical entities are introduced; the DP-GMM latent category index is a standard statistical construct.

free parameters (2)
  • DP concentration α = 10^-9 (global constant)
    Controls the birth prior P(new)=α/(α+N). Set to a single constant across datasets; chosen with knowledge of benchmark sensitivity (Tab. 5). No held-out validation split is reported for this choice.
  • n_cap (prior pseudo-count cap) = 50
    Caps the NIW degrees-of-freedom heuristic n0=min(n̄/2,ncap). Selected from ablations on CIFAR100/ImageNet100 (App. L, Tab. 15); affects birth-predictive peakiness and tail thickness.
axioms (7)
  • domain assumption Known-class feature statistics are representative of novel-class feature distributions; the shared NIW prior estimated from D_S is a valid base measure for the DP birth hypothesis.
    Sec. 3.4 uses D_S to set θ0; Eq. (12) scores birth with this θ0. App. M.3 shows degraded but still competitive performance under domain shift, confirming this is an assumption rather than a guarantee.
  • domain assumption Each category's features are approximately Gaussian with full covariance, so the NIW/Student-t predictive is a good likelihood for assign-versus-birth decisions.
    Sec. 3.3 and App. M.2 acknowledge a single full-covariance Gaussian is a first-order approximation; some categories show non-elliptical tails or substructure.
  • domain assumption DINO ViT-B/16 features fine-tuned with CE on labeled support provide a Euclidean space where category geometry is sufficiently separable for Gaussian modeling.
    Sec. 4.1; all probabilistic decisions operate in a frozen feature space. This is an empirical property, not proven.
  • standard math Blackwell-MacQueen Polya-urn predictive (Eqs. 5-6) is a valid posterior predictive for category indicators given the DP prior.
    Appendix E; standard result from Ferguson (1973) and Blackwell & MacQueen (1973).
  • standard math Gaussian-NIW conjugacy and closed-form Student-t marginals used in Eqs. 3-4 are correct.
    Appendices C-D; standard Bayesian results.
  • ad hoc to paper The heuristic n0=min(n̄/2,ncap) is a reasonable empirical-Bayes prior-strength rule.
    Eq. (11) in Sec. 3.4; justified by sensitivity (Tab. 4), not derived from first principles. Affects the birth threshold.
  • domain assumption The trace-based moment matching for κ0 is well-posed (positive solution) on all benchmarks.
    Appendix F Eq. (90); if tr(Σmeans)/tr(Σwithin) < n̄^{-1}, the estimator yields a negative κ0. No clipping strategy is stated.

pith-pipeline@v1.3.0-alltime-deepseek · 35328 in / 17031 out tokens · 176992 ms · 2026-08-02T04:59:22.597586+00:00 · methodology

0 comments
read the original abstract

On-the-fly category discovery requires deciding for each incoming test sample whether to assign it to an existing category or spawn a new one. Existing methods typically implement this decision through matching-based heuristics, such as radius- or hash-based rules. While effective in practice, these methods usually treat category birth implicitly as a fallback when no existing category matches confidently, rather than as an explicit alternative supported by its own statistical evidence. To address this, we propose DP-BOA, a posterior-predictive decision framework based on an online Dirichlet-process Gaussian mixture model with a Normal-Inverse-Wishart prior. During training, we use labeled data to calibrate a shared NIW prior over category Gaussians and warm-start the known-category posteriors. At test time, for each incoming sample, DP-BOA compares the posterior predictive evidence for assignment to existing categories against the evidence for spawning a new category induced by the DP prior, and then updates category statistics online after the decision. The method captures anisotropic category geometry and naturally adapts decision confidence as evidence accumulates. Across standard OCD benchmarks, DP-BOA consistently outperforms strong baselines and delivers particularly strong novel-class discovery performance while maintaining competitive known-class accuracy.

Figures

Figures reproduced from arXiv: 2607.13504 by Peiyan Gu, Xuming He, Zixin Teng.

Figure 1
Figure 1. Figure 1: Overview of our proposed DP-BOA framework. [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Cluster growth on CUB [PITH_FULL_IMAGE:figures/full_fig_p014_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: One-dimensional Student-t densities t_1(z \mid \mu ,\Lambda ,\nu ) with \mu {=}0 and \Lambda {=}1 for different degrees of freedom \nu \in \{0.1,1,2,5,100\} , together with the standard normal \protect \mathcal {N}(0,1) (dashed) corresponding to the limit \nu \to \infty . As \nu increases, the distribution becomes more concentrated around \mu and its tails become lighter, eventually matching the Gaussian s… view at source ↗
Figure 4
Figure 4. Figure 4: Evidence-adaptive decision boundaries for three known DP-BOA categories on the Oxford-IIIT Pet dataset. Each row corresponds to one category that is initialized from a labeled class, and columns show checkpoints where 2%, 5%, 10%, 25%, 50%, 75%, and 100% of the category’s assigned query samples have been observed (left to right). Light gray points indicate all PCA-projected features from the dataset, color… view at source ↗
Figure 5
Figure 5. Figure 5: Evidence-adaptive decision boundaries for three novel DP-BOA categories on the Oxford-IIIT Pet dataset. These clusters contain only unlabeled query samples and have no ground-truth labels. Early contours are influenced by the global NIW prior and a small, noisy empirical scatter, so they can expand or rotate as Sk rapidly reshapes the posterior covariance Ψk. As more points are assigned, Sk stabilizes, the… view at source ↗
Figure 6
Figure 6. Figure 6: Online category geometry on CUB. We analyze the feature distributions of categories in the unlabeled query split. In (a), we report the per-category scale s(\Sigma _c)=\mathrm {tr}(\Sigma _c)/d , where d is the feature dimension. In (b), we report the per-category anisotropy \kappa (\Sigma _c)=\lambda _{\max }/\lambda _{\min } . The distributions show substantial variation in scale and clear deviations fro… view at source ↗
Figure 7
Figure 7. Figure 7: Category visualizations on CIFAR100. Each panel shows one category in its own 2D PCA plane. Blue points are the category features, and the solid / dashed curves denote the fitted Gaussian 1\sigma / 2\sigma contours, respectively. is then fitted to the projected features, and its 1\sigma and 2\sigma contours are overlaid on the scatter plot. As shown in Figs. 7 and 8, many categories exhibit a dominant cent… view at source ↗
Figure 8
Figure 8. Figure 8: Category visualizations on Oxford-IIIT Pet. [PITH_FULL_IMAGE:figures/full_fig_p048_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Temporal diagnostics on Animalia. From left to right, we show the false￾birth rate, the normalized posterior cluster-mean drift, and the normalized evidence margin over five temporal bins along the query stream. Specifically, we compute the log evidence of all existing-category assignment hy￾potheses together with the explicit birth hypothesis, and define the margin as the gap between the largest and secon… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

64 extracted references · 8 linked inside Pith

  1. [1]

    Neurocom- puting5(4-5), 185–196 (1993)

    Amari, S.i.: Backpropagation and stochastic gradient descent method. Neurocom- puting5(4-5), 185–196 (1993)

    1993

  2. [2]

    In: Findings of the Association for Computational Linguistics: ACL 2024

    An, W., Shi, W., Tian, F., Lin, H., Wang, Q., Wu, Y., Cai, M., Wang, L., Chen, Y., Zhu, H., et al.: Generalized category discovery with large language models in the loop. In: Findings of the Association for Computational Linguistics: ACL 2024. pp. 8653–8665 (2024)

    2024

  3. [3]

    Transactions on Machine Learning Research (2025)

    Banerjee, A., Biswas, S.: Language-assisted feature representation and lightweight active learning for on-the-fly category discovery. Transactions on Machine Learning Research (2025)

    2025

  4. [4]

    Bishop, C.M., Nasrabadi, N.M.: Pattern recognition and machine learning, vol. 4. Springer (2006)

    2006

  5. [5]

    The annals of statistics1(2), 353–355 (1973)

    Blackwell, D., MacQueen, J.B.: Ferguson distributions via pólya urn schemes. The annals of statistics1(2), 353–355 (1973)

    1973

  6. [6]

    In: Proceedings of the IEEE/CVF international conference on computer vision

    Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 9650–9660 (2021)

    2021

  7. [7]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Choi, S., Kang, D., Cho, M.: Contrastive mean-shift learning for generalized cate- gory discovery. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 23094–23104 (2024)

    2024

  8. [8]

    In: 2009 IEEE conference on computer vision and pattern recognition

    Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large- scale hierarchical image database. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 248–255. Ieee (2009)

    2009

  9. [9]

    In: International Conference on Artificial Intelligence and Statistics

    Dinari, O., Freifeld, O.: Sampling in dirichlet process mixture models for clus- tering streaming data. In: International Conference on Artificial Intelligence and Statistics. pp. 818–835. PMLR (2022)

    2022

  10. [10]

    arXiv preprint arXiv:2010.11929 (2020)

    Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)

    Pith/arXiv arXiv 2010

  11. [11]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Du, R., Chang, D., Liang, K., Hospedales, T., Song, Y.Z., Ma, Z.: On-the-fly cate- gory discovery. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11691–11700 (2023)

    2023

  12. [12]

    Journal of the American Statistical Association90(430), 577–588 (1995)

    Escobar, M.D., West, M.: Bayesian density estimation and inference using mix- tures. Journal of the American Statistical Association90(430), 577–588 (1995)

    1995

  13. [13]

    The annals of statistics pp

    Ferguson, T.S.: A bayesian analysis of some nonparametric problems. The annals of statistics pp. 209–230 (1973)

    1973

  14. [14]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision

    Fini, E., Sangineto, E., Lathuilière, S., Zhong, Z., Nabi, M., Ricci, E.: A unified objective for novel class discovery. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9284–9292 (2021)

    2021

  15. [15]

    Jour- nal of Mathematical Psychology56(1), 1–12 (2012)

    Gershman, S.J., Blei, D.M.: A Tutorial on Bayesian Nonparametric Models. Jour- nal of Mathematical Psychology56(1), 1–12 (2012)

    2012

  16. [16]

    SIAM Journal on Computing45(5), 1762–1792 (2016)

    Ghashami, M., Liberty, E., Phillips, J.M., Woodruff, D.P.: Frequent directions: Simple and deterministic matrix sketching. SIAM Journal on Computing45(5), 1762–1792 (2016)

    2016

  17. [17]

    IEEE Transactions on Pattern Analysis and Machine Intelligence (2021) DP-BOA 17

    Han, K., Rebuffi, S.A., Ehrhardt, S., Vedaldi, A., Zisserman, A.: Autonovel: Auto- matically discovering and learning novel visual categories. IEEE Transactions on Pattern Analysis and Machine Intelligence (2021) DP-BOA 17

    2021

  18. [18]

    Han, K., Vedaldi, A., Zisserman, A.: Learning to discover novel visual categories via deeptransferclustering.In:ProceedingsoftheIEEE/CVFinternationalconference on computer vision. pp. 8401–8409 (2019)

    2019

  19. [19]

    John Wiley & Sons, Inc

    Hartigan, J.A.: Clustering algorithms. John Wiley & Sons, Inc. (1975)

    1975

  20. [20]

    He,K.,Zhang,X.,Ren,S.,Sun,J.:Deepresiduallearningforimagerecognition.In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770–778 (2016)

    2016

  21. [21]

    In: International Conference on Learning Representations (2018)

    Hsu, Y.C., Lv, Z., Kira, Z.: Learning to cluster in order to transfer across domains and tasks. In: International Conference on Learning Representations (2018)

    2018

  22. [22]

    In: International Conference on Learning Representations (2018)

    Hsu, Y.C., Lv, Z., Schlosser, J., Odom, P., Kira, Z.: Multi-class classification with- out multi-class labels. In: International Conference on Learning Representations (2018)

    2018

  23. [23]

    In: Proceedings of the IEEE/CVF international conference on computer vision

    Jia, X., Han, K., Zhu, Y., Green, B.: Joint representation learning and novel cate- gory discovery on single-and multi-modal data. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 610–619 (2021)

    2021

  24. [24]

    In: Proceedings of the IEEE international conference on computer vision workshops

    Krause, J., Stark, M., Deng, J., Fei-Fei, L.: 3d object representations for fine- grained categorization. In: Proceedings of the IEEE international conference on computer vision workshops. pp. 554–561 (2013)

    2013

  25. [25]

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

    2009

  26. [26]

    Advances in neural information processing systems25 (2012)

    Krizhevsky, A., Sutskever, I., Hinton, G.E.: Imagenet classification with deep con- volutional neural networks. Advances in neural information processing systems25 (2012)

    2012

  27. [27]

    Naval research logistics quarterly2(1-2), 83–97 (1955)

    Kuhn, H.W.: The hungarian method for the assignment problem. Naval research logistics quarterly2(1-2), 83–97 (1955)

    1955

  28. [28]

    In: Proceedings of the 19th ACM SIGKDD international conference on Knowledge discovery and data mining

    Liberty, E.: Simple and deterministic matrix sketching. In: Proceedings of the 19th ACM SIGKDD international conference on Knowledge discovery and data mining. pp. 581–588 (2013)

    2013

  29. [29]

    In: Proceedings of the IEEE/CVF International Conference on Com- puter Vision

    Liu, X., Pu, N., Zheng, H., Li, W., Sebe, N., Zhong, Z.: Generate, refine, and encode: Leveraging synthesized novel samples for on-the-fly fine-grained category discovery. In: Proceedings of the IEEE/CVF International Conference on Com- puter Vision. pp. 1078–1087 (2025)

    2025

  30. [30]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Liu, Y., Cai, Y., Jia, Q., Qiu, B., Wang, W., Pu, N.: Novel class discovery for ultra- fine-grained visual categorization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17679–17688 (2024)

    2024

  31. [31]

    arXiv preprint arXiv:2504.04804 (2025)

    Liu, Y., Han, K.: Debgcd: Debiased learning with distribution guidance for gener- alized category discovery. arXiv preprint arXiv:2504.04804 (2025)

    Pith/arXiv arXiv 2025

  32. [32]

    In: Proceedings of the Computer Vision and Pattern Recognition Conference

    Liu, Y., He, Z., Han, K.: Hyperbolic category discovery. In: Proceedings of the Computer Vision and Pattern Recognition Conference. pp. 9891–9900 (2025)

    2025

  33. [33]

    MIT press (2012)

    Murphy, K.P.: Machine learning: a probabilistic perspective. MIT press (2012)

    2012

  34. [34]

    arXiv preprint arXiv:2305.10420 (2023)

    Ouldnoughi, R., Kuo, C.W., Kira, Z.: Clip-gcd: Simple language guided generalized category discovery. arXiv preprint arXiv:2305.10420 (2023)

    Pith/arXiv arXiv 2023

  35. [35]

    In: 2012 IEEE conference on computer vision and pattern recognition

    Parkhi, O.M., Vedaldi, A., Zisserman, A., Jawahar, C.: Cats and dogs. In: 2012 IEEE conference on computer vision and pattern recognition. pp. 3498–3505. IEEE (2012)

    2012

  36. [36]

    In: Proceedings of the IEEE/CVF international conference on computer vision

    Peng, X., Bai, Q., Xia, X., Huang, Z., Saenko, K., Wang, B.: Moment matching for multi-source domain adaptation. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 1406–1415 (2019)

    2019

  37. [37]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Pu, N., Li, W., Ji, X., Qin, Y., Sebe, N., Zhong, Z.: Federated generalized category discovery. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 28741–28750 (2024) 18 P. Gu et al

    2024

  38. [38]

    In: Solla, S.A., Leen, T.K., Müller, K.R

    Rasmussen, C.E.: The infinite gaussian mixture model. In: Solla, S.A., Leen, T.K., Müller, K.R. (eds.) Advances in Neural Information Processing Systems 12. pp. 554–560. MIT Press, Cambridge, MA, USA (2000)

    2000

  39. [39]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Ronen, M., Finder, S.E., Freifeld, O.: Deepdpm: Deep clustering with an unknown number of clusters. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9861–9870 (2022)

    2022

  40. [40]

    In: Conference on Lifelong Learning Agents

    Schaeffer, R., Liu, G.K.M., Du, Y., Linderman, S., Fiete, I.R.: Streaming inference for infinite non-stationary clustering. In: Conference on Lifelong Learning Agents. pp. 310–326. PMLR (2022)

    2022

  41. [41]

    Statistica sinica pp

    Sethuraman, J.: A constructive definition of dirichlet priors. Statistica sinica pp. 639–650 (1994)

    1994

  42. [42]

    Su, Y., Zhou, R., Huang, S., Li, X., Wang, T., Wang, Z., Xu, M.: Multimodal generalizedcategorydiscovery.In:ProceedingsoftheComputerVisionandPattern Recognition Conference. pp. 1634–1643 (2025)

    2025

  43. [43]

    arXiv preprint arXiv:1906.05372 (2019)

    Tan, K.C., Liu, Y., Ambrose, B., Tulig, M., Belongie, S.: The herbarium challenge 2019 dataset. arXiv preprint arXiv:1906.05372 (2019)

    Pith/arXiv arXiv 2019

  44. [44]

    Journal of the american statistical association101(476), 1566–1581 (2006)

    Teh, Y.W., Jordan, M.I., Beal, M.J., Blei, D.M.: Hierarchical dirichlet processes. Journal of the american statistical association101(476), 1566–1581 (2006)

    2006

  45. [45]

    In: Proceedings of the IEEE conference on computer vision and pattern recognition

    Van Horn, G., Mac Aodha, O., Song, Y., Cui, Y., Sun, C., Shepard, A., Adam, H., Perona, P., Belongie, S.: The inaturalist species classification and detection dataset. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 8769–8778 (2018)

    2018

  46. [46]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recog- nition

    Vaze, S., Han, K., Vedaldi, A., Zisserman, A.: Generalized category discovery. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recog- nition. pp. 7492–7501 (2022)

    2022

  47. [47]

    Advances in Neural Information Processing Systems36, 19962–19989 (2023)

    Vaze, S., Vedaldi, A., Zisserman, A.: No representation rules them all in category discovery. Advances in Neural Information Processing Systems36, 19962–19989 (2023)

    2023

  48. [48]

    Wah, C., Branson, S., Welinder, P., Perona, P., Belongie, S.: The caltech-ucsd birds-200-2011 dataset (2011)

    2011

  49. [49]

    In: Proceedings of the fourteenth international conference on artificial intelligence and statistics

    Wang, C., Paisley, J., Blei, D.M.: Online variational inference for the hierarchical dirichlet process. In: Proceedings of the fourteenth international conference on artificial intelligence and statistics. pp. 752–760. JMLR Workshop and Conference Proceedings (2011)

    2011

  50. [50]

    arXiv preprint arXiv:2408.04591 (2024)

    Wang, H., Vaze, S., Han, K.: Hilo: A learning framework for generalized category discovery robust to domain shifts. arXiv preprint arXiv:2408.04591 (2024)

    Pith/arXiv arXiv 2024

  51. [51]

    arXiv preprint arXiv:2403.13684 (2024)

    Wang, H., Vaze, S., Han, K.: Sptnet: An efficient alternative framework for generalized category discovery with spatial prompt tuning. arXiv preprint arXiv:2403.13684 (2024)

    Pith/arXiv arXiv 2024

  52. [52]

    arXiv preprint arXiv:2407.01930 (2024)

    Wang, Y., Chen, Z., Yang, D., Sun, Y., Qi, L.: Self-cooperation knowledge distil- lation for novel class discovery. arXiv preprint arXiv:2407.01930 (2024)

    Pith/arXiv arXiv 2024

  53. [53]

    Technometrics4(3), 419–420 (1962)

    Welford, B.P.: Note on a method for calculating corrected sums of squares and products. Technometrics4(3), 419–420 (1962)

    1962

  54. [54]

    In: Proceedings of the IEEE/CVF international conference on computer vision

    Wen, X., Zhao, B., Qi, X.: Parametric classification for generalized category discov- ery: A baseline study. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 16590–16600 (2023)

    2023

  55. [55]

    IEEE Transactions on Pattern Analysis and Machine Intelligence46(3), 1513–1529 (2022)

    Willes, J., Harrison, J., Harakeh, A., Finn, C., Pavone, M., Waslander, S.L.: Bayesian embeddings for few-shot open world recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence46(3), 1513–1529 (2022)

    2022

  56. [56]

    arXiv preprint arXiv:2407.12489 (2024) DP-BOA 19

    Xu, R., Zhang, C., Ren, H., He, X.: Dual-level adaptive self-labeling for novel class discovery in point cloud segmentation. arXiv preprint arXiv:2407.12489 (2024) DP-BOA 19

    Pith/arXiv arXiv 2024

  57. [57]

    Trans- actions on Machine Learning Research (2023)

    Zhang, C., Xu, R., He, X.: Novel class discovery for long-tailed recognition. Trans- actions on Machine Learning Research (2023)

    2023

  58. [58]

    In: Proceedings of the IEEE/CVF conference on computer vision and pat- tern recognition

    Zhang, S., Khan, S., Shen, Z., Naseer, M., Chen, G., Khan, F.S.: Promptcal: Con- trastive affinity learning via auxiliary prompts for generalized novel category dis- covery. In: Proceedings of the IEEE/CVF conference on computer vision and pat- tern recognition. pp. 3479–3488 (2023)

    2023

  59. [59]

    Advances in Neural Information Processing Systems 34(2021)

    Zhao, B., Han, K.: Novel visual category discovery with dual ranking statistics and mutual knowledge distillation. Advances in Neural Information Processing Systems 34(2021)

    2021

  60. [60]

    Advances in Neural Information Processing Systems37, 101428–101455 (2024)

    Zheng, H., Pu, N., Li, W., Sebe, N., Zhong, Z.: Prototypical hash encoding for on- the-fly fine-grained category discovery. Advances in Neural Information Processing Systems37, 101428–101455 (2024)

    2024

  61. [61]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Zhong, Z., Fini, E., Roy, S., Luo, Z., Ricci, E., Sebe, N.: Neighborhood contrastive learning for novel class discovery. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10867–10875 (2021)

    2021

  62. [62]

    category

    Zhong, Z., Zhu, L., Luo, Z., Li, S., Yang, Y., Sebe, N.: Openmix: Reviving known knowledge for discovering novel visual categories in an open world. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9462–9470 (2021) 20 P. Gu et al. Contents of the Appendix – Appendix A: Implementation Details Per-dataset statistics...

    2021

  63. [63]

    keep the predictive distributions too heavy-tailed and yield poor overall accuracy, especially on novel classes. Asncap increases, performance improves substantially and is strongest around50, with a fairly broad plateau: on CIFAR- 100, the best results lie in the range50–65, while on ImageNet-100,45–55already gives very similar performance. This suggests...

  64. [64]

    We therefore fixncap=50in all experiments as a simple, robust choice near the center of this stable high-performing region

    Larger values (e.g.,100or¯n/2) further improve known-class accuracy but sharply degrade novel-class accuracy, as overly large degrees of freedom make the Student-tpredictive almost Gaussian and discourage the birth of new categories. We therefore fixncap=50in all experiments as a simple, robust choice near the center of this stable high-performing region....