pith. sign in

arxiv: 2407.19001 · v3 · submitted 2024-07-26 · 💻 cs.CV

Effective Prompt Pool Learning for Continual Category Discovery

Fernando Julio Cendra , Xinghui Li , Kai Han This is my paper

Pith reviewed 2026-05-23 23:22 UTC · model grok-4.3

classification 💻 cs.CV
keywords continual category discoveryprompt pool learningGaussian mixture promptspart-level promptingopen-world learningcatastrophic forgettingunlabeled data streamsvision transformers
0
0 comments X

The pith

Prompt pools conditioned on Gaussian mixtures for global prototypes and on part-level pools for local regions enable label-free continual discovery of new categories from unlabeled data streams.

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

The paper introduces PromptCCD and PromptCCD++ to solve continual category discovery, where a model must identify novel classes mixed with known ones in a continuous flow of unlabeled images while avoiding forgetting of earlier concepts. PromptCCD fits a Gaussian mixture model to feature embeddings so each component acts as both a prototype and a dynamic prompt that conditions the network, supporting automatic selection and count estimation without labels. PromptCCD++ further splits the prompt pool into multiple part-specific pools that assign prompts to local image regions on the fly, producing finer representations that address the finding that category count, not data volume, limits performance. If these designs work, vision systems could maintain and expand their knowledge base in open-world settings with minimal supervision.

Core claim

The paper claims that representing prompt pools via a Gaussian mixture model over global embeddings, where each component serves as both prototype and conditioning prompt, combined with decomposition into part-level prompt pools for local regions, permits label-free prompt selection, automatic estimation of emerging category counts, and improved discovery accuracy on both generic and fine-grained benchmarks while reducing catastrophic forgetting.

What carries the argument

The Gaussian Mixture Prompt (GMP) module that fits a generative GMM to feature embeddings so each mixture component doubles as class prototype and dynamic prompt, together with the Part-level Prompting (PLP) modules that maintain separate specialized prompt pools for object parts and assign them dynamically to local regions.

If this is right

  • Category count rather than sample size is the main performance bottleneck, so finer part-level representations become necessary once category numbers grow.
  • Label-free prompt selection and on-the-fly category count estimation become possible through the generative mixture model.
  • Dynamic assignment of part-specific prompts to local regions improves discovery on fine-grained data without requiring manual part labels.
  • The combined prompt-pool designs reduce catastrophic forgetting of previously discovered categories during the continual stream.
  • The frameworks achieve better discovery performance than prior methods on both generic and fine-grained benchmarks.

Where Pith is reading between the lines

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

  • The same mixture-based prompt construction could be tested on non-image modalities if their embeddings admit stable GMM fits.
  • The finding that category count dominates sample size suggests experiments that deliberately vary the number of new classes while holding total samples fixed.
  • Part-level prompt pools might generalize to other continual tasks that currently rely on global features alone.
  • If the GMM fitting step proves sensitive to embedding quality, replacing the backbone or adding embedding regularization would be a direct next test.

Load-bearing premise

That a Gaussian mixture model fitted to unlabeled feature embeddings will produce reliable class prototypes and the correct number of new categories that can be used as effective conditioning prompts for the backbone network.

What would settle it

A test stream with known ground-truth category counts where the number of mixture components selected by the GMP module differs substantially from the true number of new categories, or where adding the PLP modules produces no accuracy gain on a fine-grained discovery benchmark.

Figures

Figures reproduced from arXiv: 2407.19001 by Fernando Julio Cendra, Kai Han, Xinghui Li.

Figure 1
Figure 1. Figure 1: Overview of the Continual Category Discovery task. In the initial stage, the model learns from labelled data, while in the subsequent stages, the model learns from a continuous data stream containing unlabelled instances from known and novel classes. Recently, vision foundation models such as [6, 38] have achieved remarkable progress and shown promise in various vision tasks, from image classification and … view at source ↗
Figure 2
Figure 2. Figure 2: Our baseline CCD framework adopts a prompt-based continual learning tech￾nique by utilizing a prompt pool module to adapt the vision foundation model for CCD. Prompt learning [50, 51] has been shown effective for supervised continual learning. With properly designed prompts, the necessity of extensive modification for the model when handling the growing data stream can be greatly reduced. However, these me… view at source ↗
Figure 3
Figure 3. Figure 3: Overview of our proposed PromptCCD framework and Gaussian Mixture Prompting (GMP) module. PromptCCD continually discovers new categories while retaining previously discovered ones by learning a dynamic GMP pool to adapt the vision foundation model for CCD. Specifically, we address CCD by making use of GMP modules to estimate the probability of input zˆi by calculating the log-likelihood and use the top-k m… view at source ↗
Figure 4
Figure 4. Figure 4: t-SNE visualization of CIFAR100 with features from our model PromptCCD w/GMP and Grow & Merge on each stage. 4.3 Model Component Analysis Top-k vs random prompts. In Tab. 8, to validate the effectiveness of using top-k prompts, we compare the results by using top-k and random-k prompts. We observe that using random-k prompts hurts the performance, as evidenced by that the performance using random-k is wors… view at source ↗
Figure 5
Figure 5. Figure 5: Performance curves depicted from Tab. 7 ablation results. C U B CIFA R10 0 Best All ACC All ACC Old ACC New ACC Improving other CCD methods with our GMP. Thanks to the great flexi￾bility of our GMP, it can serve as a plug-and-play module and be seamlessly inte￾grated with other methods [PITH_FULL_IMAGE:figures/full_fig_p013_5.png] view at source ↗
read the original abstract

This paper studies effective prompt pool learning for Continual Category Discovery (CCD), a challenging open-world setting where a model must discover novel categories from a continuous stream of unlabelled data containing both known and novel classes, while mitigating catastrophic forgetting of previously learned concepts. We introduce a series of novel prompt-pool-based frameworks for CCD, each exploring a different design of prompt pools. First, we propose PromptCCD, which focuses on global class prototypes via a Gaussian Mixture Prompt (GMP) module. GMP fits a generative Gaussian mixture model over feature embeddings, where each mixture component serves as both a class prototype and a dynamic prompt that conditions the backbone's representations. This design enables label-free prompt selection and on-the-fly estimation of the number of emerging categories. Through a systematic spectrum study, we then show that category count, rather than sample size, is the primary bottleneck for discovery performance, motivating the need for finer-grained representations. Building on this finding, we propose PromptCCD++, which focuses on object-part prototypes via Part-level Prompting (PLP) modules. PLP decomposes prompt pool into multiple, specialized part-level prompt pools. During discovery phase, these pools dynamically assign part-specific prompts to local object regions without the need for manual part annotations, enabling the model to learn object-part representations that boost category discovery. Extensive evaluations on both generic and fine-grained benchmarks, supported by comprehensive ablation studies, demonstrate the effectiveness of our framework for CCD.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper introduces PromptCCD and PromptCCD++ for Continual Category Discovery (CCD). PromptCCD uses a Gaussian Mixture Prompt (GMP) module that fits a GMM to backbone features so each component acts as both a class prototype and a dynamic conditioning prompt, enabling label-free selection and on-the-fly category-count estimation. PromptCCD++ adds Part-level Prompting (PLP) modules that decompose the prompt pool into specialized part-level pools for dynamic assignment to local regions. The authors report a spectrum study showing category count (not sample size) as the main bottleneck, plus extensive evaluations and ablations on generic and fine-grained benchmarks claiming improved discovery and forgetting mitigation.

Significance. If the empirical claims hold, the work offers a concrete prompt-pool design for open-world continual discovery that avoids manual part labels and uses generative modeling for prototype-based conditioning. The spectrum study on category count versus sample size is a useful diagnostic contribution. The approach builds directly on existing prompt learning without introducing new free parameters beyond standard GMM fitting.

major comments (2)
  1. [§3.2] §3.2 (GMP module): The central mechanism relies on the EM procedure both selecting the number of components and producing stable prototypes that correctly separate known from novel classes under non-stationary streams. No analysis is provided of component stability, merge/split behavior, or the model-selection criterion when feature distributions of known and novel classes overlap; this directly affects the label-free prompt selection and on-the-fly count estimation claims.
  2. [§4] §4 (experimental validation): The reported gains in discovery performance and forgetting mitigation rest on the assumption that the fitted GMM components remain reliable across tasks. The manuscript does not include diagnostics (e.g., component assignment accuracy or prototype drift metrics) that would confirm the GMP outputs are not simply fitting noise or merged modes; without such checks the ablation results cannot isolate the contribution of the proposed modules.
minor comments (2)
  1. [§3.3] The description of how part-specific prompts are assigned to local regions in PLP (without manual annotations) would benefit from an explicit algorithmic step or pseudocode.
  2. [§3] Notation for the prompt pools and their conditioning on the backbone should be unified across GMP and PLP descriptions to avoid ambiguity in the equations.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback on the GMP module and experimental validation. We address each major comment below and will revise the manuscript to incorporate the suggested analyses.

read point-by-point responses

  1. Referee: [§3.2] §3.2 (GMP module): The central mechanism relies on the EM procedure both selecting the number of components and producing stable prototypes that correctly separate known from novel classes under non-stationary streams. No analysis is provided of component stability, merge/split behavior, or the model-selection criterion when feature distributions of known and novel classes overlap; this directly affects the label-free prompt selection and on-the-fly count estimation claims.

    Authors: We agree that further analysis of component stability and behavior under distribution overlap would strengthen the claims. The manuscript currently emphasizes end-to-end discovery performance rather than internal GMM diagnostics. In revision we will add visualizations of component evolution across tasks, quantitative measures of merge/split events (e.g., via component overlap statistics), explicit specification of the model-selection criterion used in EM, and a targeted discussion plus controlled experiments on overlapping known/novel feature distributions to support the label-free selection mechanism. revision: yes

  2. Referee: [§4] §4 (experimental validation): The reported gains in discovery performance and forgetting mitigation rest on the assumption that the fitted GMM components remain reliable across tasks. The manuscript does not include diagnostics (e.g., component assignment accuracy or prototype drift metrics) that would confirm the GMP outputs are not simply fitting noise or merged modes; without such checks the ablation results cannot isolate the contribution of the proposed modules.

    Authors: We concur that explicit reliability diagnostics would better isolate the GMP contribution from potential noise fitting. The current ablations focus on overall accuracy and forgetting metrics. In the revised manuscript we will include component assignment accuracy evaluated in controlled settings with available ground-truth labels, prototype drift metrics (e.g., average distance between successive prototypes), and additional checks confirming that components capture distinct modes rather than merged noise. These additions will be placed in §4 alongside the existing spectrum study and ablations. revision: yes

Circularity Check

0 steps flagged

No circularity: method introduces independent design choices without reducing claims to fitted inputs or self-citations

full rationale

The paper proposes PromptCCD and PromptCCD++ as new frameworks built on prompt pools, GMP (GMM fitting for prototypes/prompts), and PLP modules. These are presented as design innovations for CCD, with performance evaluated on benchmarks and ablations. No equations, derivations, or self-citation chains are shown that make any 'prediction' or result equivalent to its inputs by construction. The GMM fitting and prompt assignment are explicit modeling choices, not tautological. The spectrum study on category count is an empirical observation, not a forced outcome. This is self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

Based solely on the abstract, the central claim rests on the domain assumption that GMM components can serve as both prototypes and prompts; no free parameters or invented entities are explicitly quantified in the provided text.

axioms (1)
  • domain assumption Gaussian mixture models fitted to feature embeddings can serve as reliable class prototypes and dynamic prompts for label-free conditioning and category count estimation.
    This is the core mechanism of the GMP module described in the abstract.

pith-pipeline@v0.9.0 · 5784 in / 1274 out tokens · 24503 ms · 2026-05-23T23:22:43.838277+00:00 · methodology

discussion (0)

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

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

63 extracted references · 63 canonical work pages · 1 internal anchor

  1. [1]

    In: ICCV (2021)

    Assran, M., Caron, M., Misra, I., Bojanowski, P., Joulin, A., Ballas, N., Rabbat, M.: Semi-supervised learning of visual features by non-parametrically predicting view assignments with support samples. In: ICCV (2021)

    work page 2021

  2. [2]

    In: NeurIPS (2019)

    Berthelot, D., Carlini, N., Goodfellow, I., Papernot, N., Oliver, A., Raffel, C.: Mixmatch: A holistic approach to semi-supervised learning. In: NeurIPS (2019)

    work page 2019

  3. [3]

    IEEE TPAMI (2022)

    Boschini,M.,Bonicelli,L.,Buzzega,P.,Porrello,A.,Calderara,S.:Class-incremental continual learning into the extended der-verse. IEEE TPAMI (2022)

    work page 2022

  4. [4]

    In: NeurIPS (2020)

    Buzzega, P., Boschini, M., Porrello, A., Abati, D., Calderara, S.: Dark experience for general continual learning: a strong, simple baseline. In: NeurIPS (2020)

    work page 2020

  5. [5]

    In: ICLR (2022)

    Cao, K., Brbić, M., Leskovec, J.: Open-world semi-supervised learning. In: ICLR (2022)

    work page 2022

  6. [6]

    In: ICCV (2021)

    Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: ICCV (2021)

    work page 2021

  7. [7]

    MIT Press (2006)

    Chapelle, O., Schölkopf, B., Zien, A.: Semi-Supervised Learning. MIT Press (2006)

    work page 2006

  8. [8]

    In: ICML (2020)

    Chen, T., Kornblith, S., Norouzi, M., Hinton, G.: A simple framework for contrastive learning of visual representations. In: ICML (2020)

    work page 2020

  9. [9]

    IEEE TPAMI (2021)

    De Lange, M., Aljundi, R., Masana, M., Parisot, S., Jia, X., Leonardis, A., Slabaugh, G., Tuytelaars, T.: A continual learning survey: Defying forgetting in classification tasks. IEEE TPAMI (2021)

    work page 2021

  10. [10]

    In: ICLR (2021)

    Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. In: ICLR (2021)

    work page 2021

  11. [11]

    In: BMVC (2022)

    Fei, Y., Zhao, Z., Yang, S., Zhao, B.: Xcon: Learning with experts for fine-grained category discovery. In: BMVC (2022)

    work page 2022

  12. [12]

    IEEE TPAMI (2006)

    Fei-Fei, L., Fergus, R., Perona, P.: One-shot learning of object categories. IEEE TPAMI (2006)

    work page 2006

  13. [13]

    In: ICCV (2021)

    Fini, E., Sangineto, E., Lathuilière, S., Zhong, Z., Nabi, M., Ricci, E.: A unified objective for novel class discovery. In: ICCV (2021)

    work page 2021

  14. [14]

    Nature (2016)

    Graves, A., Wayne, G., Reynolds, M., Harley, T., Danihelka, I., Grabska-Barwińska, A., Colmenarejo, S.G., Grefenstette, E., Ramalho, T., Agapiou, J., et al.: Hybrid computing using a neural network with dynamic external memory. Nature (2016)

    work page 2016

  15. [15]

    In: ICLR (2020)

    Han, K., Rebuffi, S.A., Ehrhardt, S., Vedaldi, A., Zisserman, A.: Automatically discovering and learning new visual categories with ranking statistics. In: ICLR (2020)

    work page 2020

  16. [16]

    IEEE TPAMI (2021)

    Han, K., Rebuffi, S.A., Ehrhardt, S., Vedaldi, A., Zisserman, A.: Autonovel: Auto- matically discovering and learning novel visual categories. IEEE TPAMI (2021)

    work page 2021

  17. [17]

    In: ICCV (2019)

    Han, K., Vedaldi, A., Zisserman, A.: Learning to discover novel visual categories via deep transfer clustering. In: ICCV (2019)

    work page 2019

  18. [18]

    TMLR (2024)

    Hao, S., Han, K., Wong, K.Y.K.: Cipr: An efficient framework with cross-instance positive relations for generalized category discovery. TMLR (2024)

    work page 2024

  19. [19]

    In: CVPR (2020)

    He, K., Fan, H., Wu, Y., Xie, S., Girshick, R.: Momentum contrast for unsupervised visual representation learning. In: CVPR (2020)

    work page 2020

  20. [20]

    In: CVPR (2016)

    He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR (2016)

    work page 2016

  21. [21]

    In: ICCV (2021)

    Huang, J., Fang, C., Chen, W., Chai, Z., Wei, X., Wei, P., Lin, L., Li, G.: Trash to treasure: harvesting ood data with cross-modal matching for open-set semi- supervised learning. In: ICCV (2021)

    work page 2021

  22. [22]

    In: ICCV (2021) 16 F

    Jia, X., Han, K., Zhu, Y., Green, B.: Joint representation learning and novel category discovery on single- and multi-modal data. In: ICCV (2021) 16 F. Cendra et al

    work page 2021

  23. [23]

    In: CVPR Workshop (2022)

    Joseph, K.J., Paul, S., Aggarwal, G., Biswas, S., Rai, P., Han, K., Balasubramanian, V.N.: Spacing loss for discovering novel categories. In: CVPR Workshop (2022)

    work page 2022

  24. [24]

    In: ECCV (2022)

    Joseph, K., Paul, S., Aggarwal, G., Biswas, S., Rai, P., Han, K., Balasubramanian, V.N.: Novel class discovery without forgetting. In: ECCV (2022)

    work page 2022

  25. [25]

    In: NeurIPS (2020)

    Khosla, P., Teterwak, P., Wang, C., Sarna, A., Tian, Y., Isola, P., Maschinot, A., Liu, C., Krishnan, D.: Supervised contrastive learning. In: NeurIPS (2020)

    work page 2020

  26. [26]

    In: ICCV (2023)

    Kim, H., Suh, S., Kim, D., Jeong, D., Cho, H., Kim, J.: Proxy anchor-based unsupervised learning for continuous generalized category discovery. In: ICCV (2023)

    work page 2023

  27. [27]

    In: ICCV workshop (2013)

    Krause, J., Stark, M., Deng, J., Fei-Fei, L.: 3d object representations for fine-grained categorization. In: ICCV workshop (2013)

    work page 2013

  28. [28]

    Master’s thesis, Department of Computer Science, University of Toronto (2009)

    Krizhevsky, A., Hinton, G.: Learning multiple layers of features from tiny images. Master’s thesis, Department of Computer Science, University of Toronto (2009)

    work page 2009

  29. [29]

    In: ICLR (2017)

    Laine, S., Aila, T.: Temporal ensembling for semi-supervised learning. In: ICLR (2017)

    work page 2017

  30. [30]

    CS 231N (2015)

    Le, Y., Yang, X.: Tiny imagenet visual recognition challenge. CS 231N (2015)

    work page 2015

  31. [31]

    In: ICML (2019)

    Li, X., Zhou, Y., Wu, T., Socher, R., Xiong, C.: Learn to grow: A continual structure learning framework for overcoming catastrophic forgetting. In: ICML (2019)

    work page 2019

  32. [32]

    IEEE TPAMI (2017)

    Li, Z., Hoiem, D.: Learning without forgetting. IEEE TPAMI (2017)

    work page 2017

  33. [33]

    In: ICPR (2024)

    Liu, M., Roy, S., Zhong, Z., Sebe, N., Ricci, E.: Large-scale pre-trained models are surprisingly strong in incremental novel class discovery. In: ICPR (2024)

    work page 2024

  34. [34]

    JMLR (2008)

    Van der Maaten, L., Hinton, G.: Visualizing data using t-sne. JMLR (2008)

    work page 2008

  35. [35]

    Fine-Grained Visual Classification of Aircraft

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

    work page internal anchor Pith review Pith/arXiv arXiv 2013

  36. [36]

    In: Psychology of learning and motivation

    McCloskey, M., Cohen, N.J.: Catastrophic interference in connectionist networks: The sequential learning problem. In: Psychology of learning and motivation. Elsevier (1989)

    work page 1989

  37. [37]

    In: NeurIPS (2018)

    Oliver, A., Odena, A., Raffel, C.A., Cubuk, E.D., Goodfellow, I.: Realistic evaluation of deep semi-supervised learning algorithms. In: NeurIPS (2018)

    work page 2018

  38. [38]

    TMLR (2023)

    Oquab, M., Darcet, T., Moutakanni, T., Vo, H., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., et al.: Dinov2: Learning robust visual features without supervision. TMLR (2023)

    work page 2023

  39. [39]

    In: CVPR (2017)

    Rebuffi, S.A., Kolesnikov, A., Sperl, G., Lampert, C.H.: icarl: Incremental classifier and representation learning. In: CVPR (2017)

    work page 2017

  40. [40]

    In: NeurIPS (2021)

    Ridnik, T., Ben-Baruch, E., Noy, A., Zelnik-Manor, L.: Imagenet-21k pretraining for the masses. In: NeurIPS (2021)

    work page 2021

  41. [41]

    In: ICLR (2021)

    Rizve, M.N., Duarte, K., Rawat, Y.S., Shah, M.: In defense of pseudo-labeling: An uncertainty-aware pseudo-label selection framework for semi-supervised learning. In: ICLR (2021)

    work page 2021

  42. [42]

    In: ECCV (2022)

    Roy, S., Liu, M., Zhong, Z., Sebe, N., Ricci, E.: Class-incremental novel class discovery. In: ECCV (2022)

    work page 2022

  43. [43]

    IJCV (2015)

    Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpa- thy, A., Khosla, A., Bernstein, M., et al.: Imagenet large scale visual recognition challenge. IJCV (2015)

    work page 2015

  44. [44]

    In: NeurIPS (2021)

    Saito, K., Kim, D., Saenko, K.: Openmatch: Open-set consistency regularization for semi-supervised learning with outliers. In: NeurIPS (2021)

    work page 2021

  45. [45]

    In: NeurIPS (2020) PromptCCD 17

    Sohn, K., Berthelot, D., Li, C.L., Zhang, Z., Carlini, N., Cubuk, E.D., Kurakin, A., Zhang, H., Raffel, C.: Fixmatch: Simplifying semi-supervised learning with consistency and confidence. In: NeurIPS (2020) PromptCCD 17

    work page 2020

  46. [46]

    In: NeurIPS (2017)

    Tarvainen, A., Valpola, H.: Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. In: NeurIPS (2017)

    work page 2017

  47. [47]

    In: CVPR (2022)

    Vaze, S., Han, K., Vedaldi, A., Zisserman, A.: Generalized category discovery. In: CVPR (2022)

    work page 2022

  48. [48]

    California Institute of Technology (2011)

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

    work page 2011

  49. [49]

    In: ICLR (2024)

    Wang, H., Vaze, S., Han, K.: Sptnet: An efficient alternative framework for general- ized category discovery with spatial prompt tuning. In: ICLR (2024)

    work page 2024

  50. [50]

    In: ECCV (2022)

    Wang, Z., Zhang, Z., Ebrahimi, S., Sun, R., Zhang, H., Lee, C.Y., Ren, X., Su, G., Perot, V., Dy, J., et al.: Dualprompt: Complementary prompting for rehearsal-free continual learning. In: ECCV (2022)

    work page 2022

  51. [51]

    In: CVPR (2022)

    Wang, Z., Zhang, Z., Lee, C.Y., Zhang, H., Sun, R., Ren, X., Su, G., Perot, V., Dy, J., Pfister, T.: Learning to prompt for continual learning. In: CVPR (2022)

    work page 2022

  52. [52]

    In: ICCV (2023)

    Wen, X., Zhao, B., Qi, X.: Parametric classification for generalized category discov- ery: A baseline study. In: ICCV (2023)

    work page 2023

  53. [53]

    In: ICCV (2023)

    Wu, Y., Chi, Z., Wang, Y., Feng, S.: Metagcd: Learning to continually learn in generalized category discovery. In: ICCV (2023)

    work page 2023

  54. [54]

    In: ECCV (2020)

    Yu, Q., Ikami, D., Irie, G., Aizawa, K.: Multi-task curriculum framework for open-set semi-supervised learning. In: ECCV (2020)

    work page 2020

  55. [55]

    In: CVPR (2023)

    Zhang, S., Khan, S., Shen, Z., Naseer, M., Chen, G., Khan, F.S.: Promptcal: Contrastive affinity learning via auxiliary prompts for generalized novel category discovery. In: CVPR (2023)

    work page 2023

  56. [56]

    In: NeurIPS (2022)

    Zhang, X., Jiang, J., Feng, Y., Wu, Z.F., Zhao, X., Wan, H., Tang, M., Jin, R., Gao, Y.: Grow and merge: A unified framework for continuous categories discovery. In: NeurIPS (2022)

    work page 2022

  57. [57]

    In: NeurIPS (2021)

    Zhao, B., Han, K.: Novel visual category discovery with dual ranking statistics and mutual knowledge distillation. In: NeurIPS (2021)

    work page 2021

  58. [58]

    In: ICCV (2023)

    Zhao, B., Mac Aodha, O.: Incremental generalized category discovery. In: ICCV (2023)

    work page 2023

  59. [59]

    In: ICCV (2023)

    Zhao, B., Wen, X., Han, K.: Learning semi-supervised gaussian mixture models for generalized category discovery. In: ICCV (2023)

    work page 2023

  60. [60]

    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: CVPR (2021) PromptCCD: Learning Gaussian Mixture Prompt Pool for Continual Category Discovery –Supplementary Material– We provide this supplementary material to further support our main paper. We begin wi...

    work page 2021

  61. [61]

    • Table B, comparison on generic datasets with DINOv2

    Comparison with known class numbers: • Table A, comparison on generic datasets with DINO. • Table B, comparison on generic datasets with DINOv2. • Table C, comparison on fine-grained datasets with DINO. • Table D, comparison on fine-grained datasets with DINOv2. • Table {E,F,G,H,I},multipleruns( 5 seeds)resultsonvariantsofPromptCCD with different prompt p...

    work page

  62. [62]

    • Table K, comparison using thek-means-based estimator in [47]

    Comparison with unknown class numbers, using DINO: • Table J, comparison using our GPC-based-estimator [59]. • Table K, comparison using thek-means-based estimator in [47]. T able A: Breakdown results of various methods for CCD leveraging pretrained DINO model on generic datasets with theknown C in each unlabelled set. Stage 1ACC(%) Stage 2ACC(%) Stage 3A...

    work page arXiv 2075

  63. [63]

    in each stage are divided following the percentages in Tab. 2. To further mimic the real-world scenario, which is characterized by an abrupt increase or decrease in the number of classes of each stage, we experiment on another 3 different class splits: (1) 4:2:2:2 – the number of the unseen classes is greater than that of the seen classes; (2) 4:3:2:1 – t...

    work page