Pith. sign in

REVIEW 5 major objections 7 minor 43 references

A Hidden Stumbling Block in Generalized Category Discovery: Distracted Attention

T0 review · 5 major / 7 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read This paper claims that distracted attention on background regions is the hidden stumbling block in generalized category discovery, and that a learned token-pruning module can remove it and lift accuracy by up to 15.4%.

desk verdict A cheap token-pruning plug-in that reliably lifts fine-grained GCD accuracy, with a tuned threshold and a transfer story that needs one direct experiment before the mechanism is fully believed. read the letter →

arxiv 2507.14315 v1 pith:KUDLBDO3 submitted 2025-07-18 cs.CV

Qiyu Xu , Zhanxuan Hu , Yu Duan , Ercheng Pei , Yonghang Tai This is my paper · ORCID
classification cs.CV
keywords generalizedcategorydiscoverydistractedattentiontokenimportancemeasurementadaptivepruningvisiontransformerfocusingunknownclassbackgroundinterference
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

Generalized Category Discovery (GCD) trains a model on a few labeled classes and then expects it to cluster unlabeled images that include entirely new classes. This paper argues that the hidden bottleneck is not classifier design but attention: on unlabeled data, vision transformers spend substantial attention on irrelevant background, which corrupts the features used for clustering. To remove that noise, the authors add a lightweight module, Attention Focusing (AF), that scores every image patch for task relevance and prunes the low-scoring tokens before the final representation is formed. Integrated with the SimGCD baseline, AF raises clustering accuracy by up to 15.4 percentage points on fine-grained datasets with cluttered backgrounds, while adding almost no inference cost. The paper explicitly notes that the gain is limited when backgrounds are already simple, and that AF does not improve feature extraction from the object itself.

What carries the argument

Attention Focusing (AF), a cascade of Token Importance Measurement (TIME) and Token Adaptive Pruning (TAP). TIME adds a learnable query vector to each ViT block except the last, computes cross-attention scores $s(Q,K)=QK^{\top}/\sqrt{D}$ against patch tokens, aggregates them into an image representation through a feed-forward network, and uses an auxiliary classifier with cross-entropy loss on labeled data to train the query; a stop-gradient keeps this auxiliary path from disturbing the backbone. TAP averages the softmax-normalized per-block scores into a multi-scale importance vector, excludes the [CLS] token, sorts patch tokens by importance, and prunes with an adaptive threshold $\tau$ so that low-scoring patch tokens are discarded. The surviving tokens, including [CLS], are average-pooled as the representation fed to any existing GCD head. The multi-scale averaging is what lets the pruning adapt to images whose informative regions appear at different scales.

What would settle it

Take a GCD split in which the unlabeled (unknown) classes have systematically different background statistics from the labeled classes, for example airplanes photographed in flight as unknown classes while labeled birds sit in trees, and run SimGCD+AF; if unknown-class accuracy does not beat SimGCD without AF, or if the tokens pruned by TAP overlap substantially with ground-truth object masks, the query-transfer assumption fails.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that distracted attention is a measurable, systematic failure in GCD: the [CLS] attention maps for unlabeled images, especially images from unknown classes, spread mass over background regions, unlike labeled images. The paper shows that this distraction is not fixed by existing GCD training objectives and that it can be removed by pruning. AF inserts a learnable query vector into nearly every ViT block; cross-attention between this query and the patch tokens produces per-token importance scores, and an auxiliary classifier trained only on labeled data shapes the query to emphasize class-discriminative patches. A cascade step, TAP, averages the multi-scale importance scores and removes the least informative patch tokens with an image-adaptive threshold, then average-pools the survivors as the image representation for any GCD head. On SimGCD the module yields up to 15.4% improvement on Stanford Cars new classes with minimal extra parameters at inference, and plugging AF into CMS, SelEx, and GET also improves accuracy, supporting the claim that the distraction problem is generic.

Load-bearing premise

The load-bearing premise is that the query vector learned on labeled images transfers to unlabeled images, including never-seen classes, so its importance scores truly identify informative patches; if it does not, pruning throws away discriminative foreground tokens and the reported gains disappear.

Editorial extensions

If this is right

  • For ViT-based GCD models, AF acts as a drop-in accuracy boost: with SimGCD it raises All-class accuracy by 8.7, 13.2, and 5.2 points on CUB, Stanford Cars, and FGVC-Aircraft, and lifts unknown-class ('New') accuracy by up to 15.4 points on Stanford Cars.
  • On datasets with simple or uniform backgrounds the gains shrink: CIFAR-10/100 and Herbarium-19 show small or mixed changes, which indicates background complexity, not class difficulty, is the primary driver of distracted-attention damage.
  • Training the query only on labeled data matters: training it on all data (labeled plus unlabeled) degrades performance on every benchmark tested, so clean supervised signal is what keeps the importance scores meaningful.
  • Adaptive pruning beats fixed-k token removal: a fixed number of pruned patches saturates or reverses gains, while TAP keeps more patches on datasets with less distracting backgrounds.
  • AF generalizes: integrating it with CMS, SelEx, and GET yields consistent improvements, especially on Stanford Cars and FGVC-Aircraft.

Reading between the lines

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

  • Beyond the paper: because the query is trained only on labeled data, a natural stress test is to split labeled and unlabeled classes by background style; if the query's importance rankings shift with background statistics, AF would need a domain-generalization step to remain useful.
  • Beyond the paper: the multi-scale score aggregation in TAP could serve as an interpretability tool, showing which patches a ViT treats as informative at each layer, and the same pruning idea may help weakly supervised localization or medical imaging where background clutter dominates.
  • Beyond the paper: the resolution findings suggest a practical recipe for low-resolution inputs - upscale moderately, then prune aggressively - which could make AF useful for small-image settings without paying full-resolution inference cost.
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

5 major / 7 minor

Summary. The paper claims that generalized category discovery (GCD) methods suffer from 'distracted attention': on unlabeled images, particularly unknown-class images, the [CLS] token attends to task-irrelevant background. To fix this, the authors propose Attention Focusing (AF), a plug-in module with two components: Token Importance Measurement (TIME), which learns a query vector Q on labeled data to score patch tokens via an auxiliary classifier, and Token Adaptive Pruning (TAP), which prunes low-scoring tokens using a per-dataset threshold and pools the remaining tokens before feeding the GCD head. The method is evaluated primarily with SimGCD on seven datasets and also with CMS, SelEx, and GET on three fine-grained datasets, reporting gains of up to +15.4 accuracy on Stanford Cars New-class accuracy.

Significance. If the mechanism works as claimed, AF would be a simple and useful plug-in for GCD methods, and the paper's public code and integrations across four base methods are strengths. The multi-scale pruning analysis and the attempt to characterize when pruning helps (fine-grained versus generic datasets) are also valuable. However, the central causal claim—that pruning is driven by a query that transfers from labeled known classes to unlabeled unknown classes—is not directly tested, and the empirical support is weakened by missing error bars, inconsistent baseline numbers across tables, and per-dataset threshold tuning without a stated selection protocol.

major comments (5)
  1. [Section 4.2, Tables 1-3] No error bars or multiple seeds are reported. Several improvements are small (e.g., Herbarium-19 All +1.5 in Table 3, CIFAR100 New -1.3 in Table 2), so without variance estimates it is impossible to distinguish genuine gains from run-to-run noise. Please report mean and standard deviation over at least three seeds for the main comparisons.
  2. [Tables 1, 5, and 6] The SimGCD baseline numbers are not consistent across tables. Table 1 reports CUB All/Old/New as 60.3/65.6/57.7, while Table 5 and Table 6 report 60.1/69.7/55.4; similar discrepancies appear for Stanford Cars (53.8/71.9/45.0 vs. 55.7/73.3/47.1) and FGVC-Aircraft (54.2/59.1/51.8 vs. 53.7/64.8/48.2). This undermines the validity of the ablation comparisons, since the 'improvement' is computed against different baselines. Please state which configuration is the official one and use it uniformly.
  3. [Section 3.3, Eq. (4) and (13)] The load-bearing assumption that the query vector Q, trained only on labeled known classes, transfers to unlabeled unknown classes is explicitly stated as a hypothesis but is never directly tested. The AF(all) experiment in Table 5 is indirect because training on all data introduces noisy labels and a different optimization target. A direct test would measure, on held-out unknown classes, whether the pruned tokens overlap with foreground object regions (e.g., using ground-truth masks or bounding boxes), or compare token-score rankings for known versus unknown classes. Without such evidence, the New-class gains could be due to a generic cropping/regularization effect rather than the proposed attention-focusing mechanism.
  4. [Appendix C.7 and Table 7] The pruning threshold tau is tuned per dataset, ranging from 1e-4 to 0.2, but no selection protocol is described. Since Figure 10 shows that performance is highly sensitive to tau, it is critical to specify how tau was chosen (e.g., on a labeled validation subset) and to report performance across the tau range. Otherwise, the reported gains may reflect overfitting to the test sets.
  5. [Appendix C.4 and Section B.2] The abstract claims 'minimal computational overhead,' but Appendix C.4 shows that the CIFAR10/100 results use 112x112 input, reducing FLOPs from 16.87G to 4.7G relative to SimGCD—a substantial protocol change rather than a minimal overhead. Moreover, Section B.2 does not state the input resolution per dataset. Please clarify the exact resolution used in each main-table result and qualify the computational-overhead claim accordingly.
minor comments (7)
  1. [Eq. (6)] The definition of the pruned token set is notationally unclear: the condition 'sum_i s_i <= tau' should be expressed in terms of a threshold t on the sorted cumulative importance, and the index i should not be reused ambiguously.
  2. [References [24] and [25]] References [24] and [25] appear to be the same paper (Dynamic Conceptional Contrastive Learning for Generalized Category Discovery) duplicated with different entries.
  3. [Appendix C.7] The text says 'demonstrated in Table 7 of Section 4.3' when referring to fixed pruning, but the fixed-pruning ablation is Table 6; the cross-reference is incorrect.
  4. [Table 10 and Section 4.2] The main text attributes the CIFAR100 New-class decrease to simple backgrounds, but Appendix C.4 shows a much larger New-class drop (77.8 to 72.4) when AF uses 224x224 input; this resolution effect should be mentioned where the CIFAR results are discussed.
  5. [Contribution 1, Section 1] The contribution claims to 'quantify the harmful effects of distracted attention,' but no quantitative attention-distraction metric is defined; the paper only provides qualitative attention visualizations and downstream accuracy differences. Please either define such a metric or soften the claim.
  6. [Table 12] The 'Time consumption' entries lack units and are ambiguous; please specify whether they are per-epoch, per-iteration, or total training time.
  7. [Figure 5 caption] The caption contains 'KnownUnKnown' without a separating space, and the figure is difficult to read at the current size; please reformat it.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: AF's query is fit on labeled known classes and New-class gains are evaluated on unknown classes; the stated transfer assumption is a generalization risk, not a circular step.

full rationale

Score 0. No circularity found. The paper's derivation is self-contained: AF's token importance query Q is learned with an auxiliary cross-entropy loss (Eq. 4) on labeled known classes, and TAP prunes tokens with low multi-scale scores (Eqs. 5-6); the final GCD loss (Eq. 7) is the baseline SimGCD loss plus the auxiliary losses. The reported gains, especially New-class gains (e.g., +15.4 on Stanford Cars New), are evaluated on unlabeled data containing classes not used to fit Q, so the headline result is not forced by construction. The main caveat, stated by the authors in Sec. 3.3 ('we hypothesize that the query vector Q, learned from labeled data, generalizes well...') and in Sec. 4.2, is that Q's transfer to unknown classes is assumed rather than proven; this is an external validity/generalization risk, not a circularity. There is no fitted parameter renamed as a prediction, no self-citation chain carrying the argument, and no uniqueness theorem imported from the authors' prior work. The method is benchmarked against external methods and reference implementations.

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

The method relies on a small set of domain assumptions about attention, generalization of the learned query, and validity of pooled token representations. The main tuned parameters are the pruning threshold tau and the balance weight lambda. No new physical or conceptual entities are introduced; the query vector is a learned model parameter.

free parameters (3)
  • Pruning threshold tau = CUB 0.2, Stanford Cars 0.01, FGVC-Aircraft 0.01, CIFAR10 0.1, CIFAR100 0.1, ImageNet-100 0.05, Herbarium-19 1e-4
    Adaptive pruning threshold in Eq. 6, chosen per dataset. Sensitivity is analyzed in Appendix C.7, and the main gains depend on it.
  • Balance weight lambda = 0.05 for all datasets
    Weights the auxiliary TIME loss in Eq. 7. Tuned with ablation in Appendix C.7.
  • Learnable query vector Q = Learned during training
    Cross-attention query in each TIME module, trained with cross-entropy on labeled data. This is the core learned parameter that determines token importance.
assumptions (4)
  • domain assumption A query vector trained only on labeled data generalizes to unlabeled data, including unknown classes
    Invoked in Section 3.3 to justify using TIME on unlabeled data. If false, pruning would harm new-class accuracy.
  • domain assumption Attention mass on background regions is a cause of GCD performance loss, and pruning those tokens improves representations
    Basis of the method, supported only by qualitative attention maps in Fig. 1 and Fig. 8, not by a quantitative distraction metric.
  • domain assumption Multi-scale importance scores from ViT blocks identify task-relevant objects
    Justifies averaging scores across blocks in Eq. 5. Ablation in Fig. 4 and Fig. 5 is qualitative.
  • domain assumption Average pooling of remaining tokens is a sufficient representation for GCD
    Backbone output is changed from the [CLS] token to pooled tokens. Table 11 supports this empirically, but it is a design choice asserted as better.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A Hidden Stumbling Block in Generalized Category Discovery: Distracted Attention." pith.science (2026). https://pith.science/paper/KUDLBDO3

@misc{pith2026250714315,
  author       = {Pith},
  title        = {Pith review of: A Hidden Stumbling Block in Generalized Category Discovery: Distracted Attention},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KUDLBDO3}},
  note         = {Machine review of arXiv:2507.14315}
}
read the original abstract

Generalized Category Discovery (GCD) aims to classify unlabeled data from both known and unknown categories by leveraging knowledge from labeled known categories. While existing methods have made notable progress, they often overlook a hidden stumbling block in GCD: distracted attention. Specifically, when processing unlabeled data, models tend to focus not only on key objects in the image but also on task-irrelevant background regions, leading to suboptimal feature extraction. To remove this stumbling block, we propose Attention Focusing (AF), an adaptive mechanism designed to sharpen the model's focus by pruning non-informative tokens. AF consists of two simple yet effective components: Token Importance Measurement (TIME) and Token Adaptive Pruning (TAP), working in a cascade. TIME quantifies token importance across multiple scales, while TAP prunes non-informative tokens by utilizing the multi-scale importance scores provided by TIME. AF is a lightweight, plug-and-play module that integrates seamlessly into existing GCD methods with minimal computational overhead. When incorporated into one prominent GCD method, SimGCD, AF achieves up to 15.4% performance improvement over the baseline with minimal computational overhead. The implementation code is provided in https://github.com/Afleve/AFGCD.

Figures

Figures reproduced from arXiv: 2507.14315 by the authors.

Figure 1
Figure 1. The masks obtained by thresholding the self-attention [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. The pipeline of GCD with our proposed Attention Fo￾cusing(AF) mechanism. AF consists of two components: Token Importance Measurement (TIME) and Token Adaptive Pruning (TAP), working in a cascade. Here, the ’Head’ can be inherited from any existing GCD model. paramount importance. Additionally, many methods rely on the [CLS] token for pruning, but in the GCD task, the [CLS] token for unlabeled data tends to be of low… view at source ↗
Figure 3
Figure 3. The internal pipeline of TIME. The red dashed lines rep [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: Investigation of Multi-scale token importance measurement. ”SimGCD+AF-” refers to a setting where only the query from the [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: The results of token pruning using query vectors from each layer. Specifically, the last column illustrates the multi-scale token [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: The dynamic change of the number of retaining patches [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: Image examples from the used datasets. in [PITH_FULL_IMAGE:figures/full_fig_p012_7.png]
Figure 8
Figure 8. Figure 8: The masks obtained by thresholding the self-attention maps to retain same percent of the total mass cross different methods. [PITH_FULL_IMAGE:figures/full_fig_p013_8.png]
Figure 9
Figure 9. Figure 9: The partitions of input images with the same patch size under different resolutions. [PITH_FULL_IMAGE:figures/full_fig_p014_9.png]
Figure 10
Figure 10. Figure 10: Investigation of the parameter λ and τ . between retaining relevant features and eliminating redun￾dancy. Despite these variations, AF consistently achieves significant performance improvements across different λ, demonstrating its robustness and effectiveness in enha…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

43 extracted references · 40 canonical work pages

  1. [1]

    Iterative patch selection for high-resolution image recognition

    Benjamin Bergner, Christoph Lippert, and Aravindh Ma- hendran. Iterative patch selection for high-resolution image recognition. In International Conference on Learning Rep- resentations, 2022. 3

    work page 2022

  2. [2]

    Token cropr: Faster vits for quite a few tasks

    Benjamin Bergner, Christoph Lippert, and Aravindh Mahen- dran. Token cropr: Faster vits for quite a few tasks. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 9740–9750, 2025. 3, 4

    work page 2025

  3. [3]

    Token merging: Your ViT but faster

    Daniel Bolya, Cheng-Yang Fu, Xiaoliang Dai, Peizhao Zhang, Christoph Feichtenhofer, and Judy Hoffman. Token merging: Your ViT but faster. In International Conference on Learning Representations, 2023. 3

    work page 2023

  4. [4]

    Emerg- ing properties in self-supervised vision transformers

    Mathilde Caron, Hugo Touvron, Ishan Misra, Herv ´e J´egou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerg- ing properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9650–9660, 2021. 12

    work page 2021

  5. [5]

    Parametric information max- imization for generalized category discovery

    Florent Chiaroni, Jose Dolz, Ziko Imtiaz Masud, Amar Mitiche, and Ismail Ben Ayed. Parametric information max- imization for generalized category discovery. InProceedings of the IEEE/CVF International Conference on Computer Vi- sion, pages 1729–1739, 2023. 6, 7

    work page 2023

  6. [6]

    Contrastive mean- shift learning for generalized category discovery

    Sua Choi, Dahyun Kang, and Minsu Cho. Contrastive mean- shift learning for generalized category discovery. In Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2, 4, 6, 7

    work page 2024

  7. [7]

    Imagenet: A large-scale hierarchical image database

    Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 248–255. Ieee, 2009. 5, 12

    work page 2009

  8. [8]

    An image is worth 16x16 words: Transformers for image recognition at scale

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Matthias Minderer Mostafa Dehghani, Georg Heigold, Syl- vain Gelly, Jakob Uszkoreit, , and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 12

    work page 2021

Show all 43 references
  1. [9]

    A unified objective for novel class discovery

    Enrico Fini, Enver Sangineto, St ´ephane Lathuili `ere, Zhun Zhong, Moin Nabi, and Elisa Ricci. A unified objective for novel class discovery. InProceedings of the IEEE/CVF Inter- national Conference on Computer Vision, pages 9284–9292,

  2. [10]

    Semi-supervised learning made simple with self-supervised clustering

    Enrico Fini, Pietro Astolfi, Karteek Alahari, Xavier Alameda-Pineda, Julien Mairal, Moin Nabi, and Elisa Ricci. Semi-supervised learning made simple with self-supervised clustering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 31...

    2023

  3. [11]

    LLaV A-UHD: an lmm perceiving any aspect ratio and high- resolution images

    Zonghao Guo, Ruyi Xu, Yuan Yao, Junbo Cui, Zanlin Ni, Chunjiang Ge, Tat-Seng Chua, Zhiyuan Liu, and Gao Huang. LLaV A-UHD: an lmm perceiving any aspect ratio and high- resolution images. In ECCV, 2024. 3

    2024

  4. [12]

    Learning to discover novel visual categories via deep transfer cluster- ing

    Kai Han, Andrea Vedaldi, and Andrew Zisserman. Learning to discover novel visual categories via deep transfer cluster- ing. In Proceedings of the IEEE/CVF International Confer- ence on Computer Vision, pages 8401–8409, 2019. 6

    2019

  5. [13]

    Autonovel: Automati- cally discovering and learning novel visual categories

    Kai Han, Sylvestre-Alvise Rebuffi, Sebastien Ehrhardt, An- drea Vedaldi, and Andrew Zisserman. Autonovel: Automati- cally discovering and learning novel visual categories. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(10):6767–6781, 2021. 6

    2021

  6. [14]

    Spvit: Enabling faster vision transformers via latency-aware soft token pruning

    Zhenglun Kong, Peiyan Dong, Xiaolong Ma, Xin Meng, Wei Niu, Mengshu Sun, Xuan Shen, Geng Yuan, Bin Ren, Hao Tang, et al. Spvit: Enabling faster vision transformers via latency-aware soft token pruning. In Computer Vision– ECCV 2022: 17th European Conference, Tel Aviv, Israel, ...

    2022

  7. [15]

    3d object representations for fine-grained categorization

    Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops, pages 554–561, 2013. 5, 11, 12

    2013

  8. [16]

    Learning multiple layers of features from tiny images

    Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 5, 12

    2009

  9. [17]

    The hungarian method for the assignment problem

    Harold W Kuhn. The hungarian method for the assignment problem. In Naval research logistics quarterly, 1955. 5

    1955

  10. [18]

    Comatch: Semi-supervised learning with contrastive graph regulariza- tion

    Junnan Li, Caiming Xiong, and Steven CH Hoi. Comatch: Semi-supervised learning with contrastive graph regulariza- tion. In Proceedings of the IEEE/CVF International Confer- ence on Computer Vision, pages 9475–9484, 2021. 2

    2021

  11. [19]

    Mon- key: Image resolution and text label are important things for large multi-modal models

    Zhang Li, Biao Yang, Qiang Liu, Zhiyin Ma, Shuo Zhang, Jingxu Yang, Yabo Sun, Yuliang Liu, and Xiang Bai. Mon- key: Image resolution and text label are important things for large multi-modal models. In proceedings of the IEEE/CVF conference on computer vision and pattern recognition ,

  12. [20]

    Not all patches are what you need: Expediting vision transformers via token reorganiza- tions

    Youwei Liang, Chongjian Ge, Zhan Tong, Yibing Song, Jue Wang, and Pengtao Xie. Not all patches are what you need: Expediting vision transformers via token reorganiza- tions. In International Conference on Learning Representa- tions, 2022. 3

    2022

  13. [21]

    Revisiting token pruning for object detection and instance segmentation

    Yifei Liu, Mathias Gehrig, Nico Messikommer, Marco Can- nici, and Davide Scaramuzza. Revisiting token pruning for object detection and instance segmentation. In Proceed- ings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), 2024. 3

    2024

  14. [22]

    Fine-grained visual classi- fication of aircraft

    Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew Blaschko, and Andrea Vedaldi. Fine-grained visual classi- fication of aircraft. arXiv preprint arXiv:1306.5151 , 2013. 5, 11, 12

    2013 arXiv

  15. [23]

    Mos: Modeling object-scene associations in generalized category discovery

    Zhengyuan Peng, Jinpeng Ma, Zhimin Sun, Ran Yi, Haichuan Song, Xin Tan, and Lizhuang Ma. Mos: Modeling object-scene associations in generalized category discovery. In Proceedings of the Computer Vision and Pattern Recogni- tion Conference, pages 15118–15128, 2025. 3, 6 9

    2025

  16. [24]

    Dynamic conceptional contrastive learning for generalized category discovery

    Nan Pu, Zhun Zhong, and Nicu Sebe. Dynamic conceptional contrastive learning for generalized category discovery. In Proceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition, pages 7579–7588, 2023. 6

    2023

  17. [25]

    Dynamic conceptional contrastive learning for generalized category discovery

    Nan Pu, Zhun Zhong, and Nicu Sebe. Dynamic conceptional contrastive learning for generalized category discovery. In Proceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition, pages 7579–7588, 2023. 2

    2023

  18. [26]

    Learn to categorize or categorize to learn? self-coding for general- ized category discovery

    Sarah Rastegar, Hazel Doughty, and Cees Snoek. Learn to categorize or categorize to learn? self-coding for general- ized category discovery. In Advances in Neural Information Processing Systems, 2023. 2, 6, 7

    2023

  19. [27]

    Selex: Self-expertise in fine-grained generalized category discovery

    Sarah Rastegar, Mohammadreza Salehi, Yuki M Asano, Hazel Doughty, and Cees GM Snoek. Selex: Self-expertise in fine-grained generalized category discovery. In European Conference on Computer Vision , pages 440–458. Springer,

  20. [28]

    Open-set recognition: A good closed-set classifier is all you need? In arXiv preprint arXiv:2110.06207, 2021

    Andrea Vedaldi Sagar Vaze, Kai Han and Andrew Zisser- man. Open-set recognition: A good closed-set classifier is all you need? In arXiv preprint arXiv:2110.06207, 2021. 5, 11

    2021 arXiv

  21. [29]

    The herbarium challenge 2019 dataset

    Kiat Chuan Tan, Yulong Liu, Barbara Ambrose, Melissa Tulig, and Serge Belongie. The herbarium challenge 2019 dataset. arXiv preprint arXiv:1906.05372, 2019. 5, 12

    2019 arXiv

  22. [30]

    Patch slimming for ef- ficient vision transformers

    Yehui Tang, Kai Han, Yunhe Wang, Chang Xu, Jianyuan Guo, Chao Xu, and Dacheng Tao. Patch slimming for ef- ficient vision transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 12165–12174, 2022. 3

    2022

  23. [31]

    Generalized category discovery

    Sagar Vaze, Kai Han, Andrea Vedaldi, and Andrew Zisser- man. Generalized category discovery. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7492–7501, 2022. 2, 6, 7

    2022

  24. [32]

    Get: Unlocking the multi-modal potential of clip for generalized category dis- covery

    Enguang Wang, Zhimao Peng, Zhengyuan Xie, Fei Yang, Xialei Liu, and Ming-Ming Cheng. Get: Unlocking the multi-modal potential of clip for generalized category dis- covery. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 20296–20306, 2025. 2, 4, 6, 7

    2025

  25. [33]

    Sptnet: An effi- cient alternative framework for generalized category discov- ery with spatial prompt tuning

    Hongjun Wang, Sagar Vaze, and Kai Han. Sptnet: An effi- cient alternative framework for generalized category discov- ery with spatial prompt tuning. In International Conference on Learning Representations (ICLR), 2024. 2, 6, 7

    2024

  26. [34]

    Caltech-ucsd birds 200

    Peter Welinder, Steve Branson, Takeshi Mita, Catherine Wah, Florian Schroff, Serge Belongie, and Pietro Perona. Caltech-ucsd birds 200. 2010. 2, 5, 11, 12

    2010

  27. [35]

    Parametric classification for generalized category discovery: A baseline study

    Xin Wen, Bingchen Zhao, and Xiaojuan Qi. Parametric classification for generalized category discovery: A baseline study. In Proceedings of the IEEE/CVF International Con- ference on Computer Vision, pages 16590–16600, 2023. 1, 2, 4, 5, 6, 7, 12, 14

    2023

  28. [36]

    Llama-adapter: Efficient finetuning of language models with zero-init attention

    Renrui Zhang, Jiaming Han, Chris Liu, Peng Gao, Ao- jun Zhou, Xiangfei Hu, Shilin Yan, Pan Lu, Hongsheng Li, and Yu Qiao. Llama-adapter: Efficient finetuning of language models with zero-init attention. arXiv preprint arXiv:2303.16199, 2023. 3

    2023 arXiv

  29. [37]

    Prompt- cal: Contrastive affinity learning via auxiliary prompts for generalized novel category discovery

    Sheng Zhang, Salman Khan, Zhiqiang Shen, Muzammal Naseer, Guangyi Chen, and Fahad Shahbaz Khan. Prompt- cal: Contrastive affinity learning via auxiliary prompts for generalized novel category discovery. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern R...

    2023

  30. [38]

    Less attention is more: Prompt trans- former for generalized category discovery

    Wei Zhang, Baopeng Zhang, Zhu Teng, Wenxin Luo, Junnan Zou, and Jianping Fan. Less attention is more: Prompt trans- former for generalized category discovery. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 30322–30331, 2025. 3, 6

    2025

  31. [39]

    Learning semi- supervised gaussian mixture models for generalized category discovery

    Bingchen Zhao, Xin Wen, and Kai Han. Learning semi- supervised gaussian mixture models for generalized category discovery. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16623–16633, 2023. 2

    2023

  32. [40]

    Learning semi- supervised gaussian mixture models for generalized cate- gory discovery

    Bingchen Zhao, Xin Wen, and Kai Han. Learning semi- supervised gaussian mixture models for generalized cate- gory discovery. In Proceedings of the IEEE/CVF Inter- national Conference on Computer Vision (ICCV) , pages 16623–16633, 2023. 2, 6 10 A Hidden Stumbling Block in Gener...

    2023

  33. [41]

    Classifier Learning aims to train a classifier that assigns labels to unlabeled data. Within the SimGCD framework, this objective is achieved through a paramet- ric classifier refined via a self-distillation strategy, where the number of categories, denoted as |Yu|, is predete...

  34. [42]

    As shown in Figure 10, it is evident that τ can yield significant performance improvements within a specific range

    Hyperparameter τ For τ, we maintain λ = 0 .05, while varying τ with a same interval. As shown in Figure 10, it is evident that τ can yield significant performance improvements within a specific range. However, the influence of τ on model per- formance is particularly pronounce...

  35. [43]

    As shown in Figure 10, it can be ob- served that the performance of AF declines whenλ ≤ 0.03

    Hyperparameter λ For λ, we maintain τ as the pre-set value for the corre- sponding dataset, while varying within the set λ = {0.01, 0.03, 0.05, 0.07, 0.1}. As shown in Figure 10, it can be ob- served that the performance of AF declines whenλ ≤ 0.03. We attribute this phenomeno...

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.