Pith. sign in

REVIEW 3 major objections 6 minor 73 references

FTCFormer: Fuzzy Token Clustering Transformer for Image Classification

T0 review · 3 major / 6 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 FTCFormer claims that semantic token clustering via DPC-FKNN, SCS, and channel-level merging outperforms TCFormer on all 32 tested image-classification datasets.

desk verdict Plausible TCFormer refinement with broad but noisy evidence, though a direct arithmetic contradiction in the main ablation table undermines the component attribution. read the letter →

arxiv 2507.10283 v1 pith:YNQMOVCH submitted 2025-07-14 cs.CV

Muyi Bao , Changyu Zeng , Yifan Wang , Zhengni Yang , Zimu Wang , Guangliang Cheng , Jun Qi , Wei Wang This is my paper · ORCID
classification cs.CV
keywords FTCFormerFuzzyTokenClusteringDPC-FKNNSpatialConnectivityScoreChannelMergingVisionTransformerImageClassification
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

FTCFormer is a vision transformer that forms its visual tokens by semantic clustering instead of a fixed spatial grid, so it spends more tokens on informative regions like faces or lesions and fewer on background. The paper's central claim is that this approach, implemented in the FTCM downsampling module, beats the previous token-clustering transformer TCFormer on every one of 32 datasets spanning fine-grained, natural, medical, remote sensing, MNIST-like, and CIFAR-like images. The reported average gains range from 0.06% on CIFAR-like data to 1.43% on fine-grained data, with only modest added cost (0.26 GFLOPs and 0.38M parameters). If these results hold, semantic clustering tokenization is a reliable, cheap upgrade for hierarchical image classifiers.

What carries the argument

The FTCM (Fuzzy Token Clustering and Merging) module is the central object. It is a downsampling stage that replaces grid-based pooling between transformer stages. Its operation rests on three named components: DPC-FKNN, a density-peak clustering rule with a fuzzy K-nearest-neighbor kernel; SCS, a spatial connectivity score assigning non-center tokens to clusters; and Cmerge, a channel-wise merging formula. The mathematical anchor is the fuzzy distance kernel µ(i,j) defined in Eq. (1), which gives exponentially higher weight to closer pairs and extra weight to KNN members, with a distance-decay factor φ applied to all other pairs. The module also uses cross-attention between merged and original tokens.

What would settle it

Retrain FTCFormer on any fine-grained dataset from the paper while replacing the DPC-FKNN center-selection step with random token centers (keeping SCS and Cmerge unchanged); if accuracy does not drop well below the reported ablation value, the clustering-center mechanism is not the source of the gains. Alternatively, grid-search φ in Eq. (1) from small to large values on Flowers102 and Stanford Cars; if the optimal φ shifts with the dataset, the fixed form of the fuzzy kernel is not as universal as the paper assumes.

Watch

Extended reading notes

Core claim

The central claim is that FTCFormer outperforms TCFormer on all 32 datasets tested, with average accuracy improvements of 1.43% on fine-grained, 1.09% on natural, 0.97% on medical, 0.55% on remote sensing, 0.21% on MNIST-like, and 0.06% on CIFAR-like datasets. The paper attributes the improvement to three mechanisms inside the FTCM module: DPC-FKNN for cluster-center selection, which weights both the K-nearest neighbors and distance-decayed non-neighbors when estimating local density; SCS for token assignment, which combines shared-neighbor count with neighbor closeness and falls back to Euclidean distance only when SCS is zero; and Cmerge for token merging, which computes importance scores per channel rather than per token. Ablation experiments show each mechanism contributes, with DPC-FKNN delivering the largest single gain (e.g. +2.89% on Stanford Cars). The authors also report that FTCM adds only 0.26 GFLOPs and 0.38M parameters over the TCFormer baseline, and they perform hyperparameter sweeps showing the default values (KFuzzy=KSCS=5) are near-optimal across several datasets.

Load-bearing premise

The load-bearing assumption is that the fuzzy distance kernel, the shared-neighbor connectivity score, and the channel-wise merging weights all measure the right notion of semantic similarity inside the transformer's learned feature maps; if these hand-designed heuristics do not align with what actually matters for classification, the reported gains will not hold.

Editorial extensions

If this is right

  • If the reported 32-dataset comparison is accurate, semantic clustering downsampling can be used as a drop-in replacement for grid-based downsampling in hierarchical transformers, improving accuracy without changing the downstream head.
  • The largest gains on fine-grained datasets suggest the module helps most where classes differ by subtle local detail, such as flower species, car models, and aircraft types.
  • Because the gains persist on medical and remote sensing data, the method may transfer to tasks where region importance is highly skewed, such as pathology slides or aerial scenes.
  • The small added cost implies the improvement is not simply a matter of more parameters or compute; the clustering logic itself is doing the work.
  • The saturation behavior on MNIST-like and CIFAR-like sets shows the method remains beneficial even when the baseline is already near 99% accuracy.

Reading between the lines

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

  • The paper never specifies how the standard-deviation parameter φ in Eq. (1) is set, even though it controls the kernel's decay for non-neighbor tokens; a natural extension would be to test a learned or data-dependent φ to see if the gains change.
  • The same semantic clustering machinery could be ported to other dense prediction tasks, where allocating more tokens to object regions may improve segmentation or detection as well as classification.
  • The ablation isolates DPC-FKNN as the main source of gain, so a minimal reproduction experiment—swapping only the center-selection step back to DPC-KNN—would directly test whether fuzzy neighbor weighting is the essential ingredient.
  • A stress test on heavily class-imbalanced datasets would reveal whether the method's bias toward dense, semantically important regions helps rare classes or simply concentrates tokens on the majority class.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The paper proposes FTCFormer, a hierarchical vision transformer that replaces grid-based downsampling with a fuzzy token clustering and merging (FTCM) module. The module contains three components: DPC-FKNN for clustering-center selection, a Spatial Connectivity Score (SCS) for token assignment, and channel-wise merging (Cmerge) for token fusion. The authors report experiments on 32 image-classification datasets across seven domains, claiming consistent accuracy improvements over the TCFormer baseline, with average gains of 1.43% on fine-grained, 1.09% on natural, 0.97% on medical, 0.55% on remote sensing, 0.21% on MNIST-like, and 0.06% on CIFAR-like datasets. Additional ablations compare FTCM with grid-based downsampling and isolate the contribution of each component.

Significance. If the empirical claims hold, the paper makes a useful contribution by showing that semantic token clustering can be made more robust through fuzzy KNN, spatial connectivity, and channel-level merging, and by validating this across a broad set of datasets. The release of code and the breadth of the evaluation are strengths, as is the direct comparison with TCFormer under the same architecture. However, the central claim of consistent improvement is currently undermined by an internal inconsistency in the component ablation, an under-specified kernel parameter, and the absence of statistical uncertainty estimates. The paper's contribution remains plausible, but the evidence as presented is not yet reliable enough for acceptance.

major comments (3)
  1. [Table 5] The Stanford Cars column of Table 5 is internally inconsistent. With baseline 81.83, DPC-FKNN gives 84.72 (+2.89), adding SCS gives 85.33 (+0.61), and adding Cmerge is printed as 84.55 (+0.22), with a total improvement of +2.72. Under the cumulative reading used in the other columns, the full model should be 85.55 after the Cmerge increment of +0.22. If 84.55 is the true full-model accuracy (it is also the value reported in Table 2a), then the Cmerge contribution is actually -0.78, contradicting the printed +0.22 and the text's claim that Cmerge 'further enhances performance.' Since Table 5 is the direct evidence for the component-wise attribution that underlies the paper's central claim, this discrepancy must be resolved before the contribution of Cmerge and the interpretation of the gains can be accepted.
  2. [Eq. (1), Section 3.2.1] The fuzzy distance kernel in Eq. (1) is under-defined because the parameter φ is never specified. The text says φ is 'the standard deviation representing the sparsity of all tokens,' but no formula is given for computing it, and it is not stated whether it is a scalar, a per-token value, or something else. Since this kernel directly defines the local density ρ_i in Eq. (2) and the DPC-FKNN clustering-center selection, the method is not reproducible without this definition. Please provide the exact computation of φ and clarify its role in the non-KNN term.
  3. [Section 4.5.3, Figure 4, Section 4.2] The central empirical claim of 'consistent improvements' over TCFormer is weakened by the absence of error bars or repeated-run statistics. Many reported gains are below 1% (for example, 0.06% on CIFAR-like and 0.21% on MNIST-like datasets), and no standard deviations or number of trials are reported, so it is unclear whether these differences exceed training variance. In addition, the hyperparameters K_Fuzzy and K_SCS are tuned on CIFAR100, DTD, and Flowers102 (Section 4.5.3 and Figure 4), and these same datasets appear in the main results with the selected values K_Fuzzy=K_SCS=5. This selection on the evaluation datasets makes the reported improvements on those datasets optimistic. Please report uncertainty estimates and either use a held-out validation split for hyperparameter selection or discuss the resulting selection bias quantitatively.
minor comments (6)
  1. [Abstract and Section 1] The sentence 'achieving gains of improving 1.43%' is ungrammatical; it should be 'achieving average improvements of 1.43%'.
  2. [Table 2] Table 2a has the typo 'Standford Cars' and Table 2b has 'RESUS45'; these should read 'Stanford Cars' and 'RESISC45', respectively.
  3. [Figure 3 caption] The caption contains a stray comma and period: 'across different datasets. , including natural' should be cleaned up.
  4. [Eq. (4)] The definition of SNN(i,j) is written as a set of pairs but used as a count |SNN(i,j)|. Please clarify that it denotes the intersection of the KNN sets, and avoid the confusing set-builder notation.
  5. [Section 4.5.1 and Table 4] The text states 'FTCM consistently outperforms FTM,' but the baseline in Table 4 is labeled TCM. Please make the naming consistent and ensure the text refers to the same method.
  6. [Eq. (7)] The channel-level importance score P is not described beyond the claim that it is regressed by a linear layer. Please state how P is obtained and how it is trained, since this is central to the Cmerge mechanism.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: FTCFormer's claimed gains are empirical comparisons, not consequences of its definitions, though Table 5 contains an internal arithmetic inconsistency.

full rationale

The paper's derivation chain is not circular. The FTCM module is specified concretely in Eqs. (1)-(8) as an architectural proposal, and the central claim is an empirical comparison against TCFormer on held-out validation splits of 32 datasets. No equation is defined in terms of the target accuracy, and no fitted parameter is renamed as a prediction: K_Fuzzy and K_SCS are fixed at 5 for all datasets, and the reported accuracy numbers are measurements rather than algebraic consequences of the hyperparameter choice. The hyperparameter ablation in Sec. 4.5.3 uses CIFAR100, DTD, and Flowers102, which also appear in the main results, so there is a mild selection-on-validation concern, but this does not make the reported accuracies equivalent to the model's inputs by construction. Citations to TCFormer are to external prior work, not to the present authors, and no uniqueness theorem or ansatz is smuggled in via self-citation. The most notable reviewable issue is an internal arithmetic inconsistency in Table 5: on Stanford Cars, 81.83 + 2.89 + 0.61 + 0.22 sums to 85.55, while the printed full-model accuracy is 84.55 and the printed total improvement is +2.72. This inconsistency undermines the component-wise attribution claim but is a correctness and reproducibility problem, not circularity. The paper is therefore self-contained with respect to the circularity criteria.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The central claim rests on the DPC heuristic, the proposed kernel and metrics, and the fairness of the evaluation protocol; none of these are derived from first principles, so the contribution is empirical rather than theoretical.

free parameters (3)
  • K_Fuzzy = 5
    K value for DPC-FKNN local density in Eq. (2); chosen after ablation on CIFAR100, DTD, Flowers102 and fixed for all datasets.
  • K_SCS = 5
    K value for SCS metric in Eq. (6); ablation shows best near 5, fixed for all datasets.
  • phi (φ) = not specified
    Standard deviation in Eq. (1) controlling attenuation for non-KNN neighbors; the paper provides no formula or estimation method, leaving the kernel behavior under-specified.
assumptions (5)
  • domain assumption Density peaks clustering assumption: cluster centers should have both higher local density and larger separation distance.
    Used to select centers by maximizing rho_i * delta_i in Section 3.2.1; this is an unproved heuristic from the clustering literature.
  • ad hoc to paper The fuzzy distance kernel mu(i,j) in Eq. (1) is an appropriate similarity measure for learned feature tokens.
    Introduced without theoretical justification; its success is an empirical premise.
  • ad hoc to paper Channel-level importance scores better preserve discriminative information than token-level scores.
    Underlies Cmerge in Eq. (7); supported only by the ablation in Table 5.
  • ad hoc to paper The SCS metric captures meaningful spatial connectivity in high-dimensional feature spaces.
    Proposed in Section 3.2.2 as a remedy for Euclidean distance; no formal guarantee, only empirical support.
  • domain assumption Random 8:2 train/validation splits yield results comparable to official dataset partitions where official splits exist.
    Section 4.1 applies random splits only in the absence of official partitions; the fairness of cross-paper comparisons depends on this assumption.

how reviews work

0 comments
Cite this review

Pith. "Pith review of FTCFormer: Fuzzy Token Clustering Transformer for Image Classification." pith.science (2026). https://pith.science/paper/YNQMOVCH

@misc{pith2026250710283,
  author       = {Pith},
  title        = {Pith review of: FTCFormer: Fuzzy Token Clustering Transformer for Image Classification},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YNQMOVCH}},
  note         = {Machine review of arXiv:2507.10283}
}
read the original abstract

Transformer-based deep neural networks have achieved remarkable success across various computer vision tasks, largely attributed to their long-range self-attention mechanism and scalability. However, most transformer architectures embed images into uniform, grid-based vision tokens, neglecting the underlying semantic meanings of image regions, resulting in suboptimal feature representations. To address this issue, we propose Fuzzy Token Clustering Transformer (FTCFormer), which incorporates a novel clustering-based downsampling module to dynamically generate vision tokens based on the semantic meanings instead of spatial positions. It allocates fewer tokens to less informative regions and more to represent semantically important regions, regardless of their spatial adjacency or shape irregularity. To further enhance feature extraction and representation, we propose a Density Peak Clustering-Fuzzy K-Nearest Neighbor (DPC-FKNN) mechanism for clustering center determination, a Spatial Connectivity Score (SCS) for token assignment, and a channel-wise merging (Cmerge) strategy for token merging. Extensive experiments on 32 datasets across diverse domains validate the effectiveness of FTCFormer on image classification, showing consistent improvements over the TCFormer baseline, achieving gains of improving 1.43% on five fine-grained datasets, 1.09% on six natural image datasets, 0.97% on three medical datasets and 0.55% on four remote sensing datasets. The code is available at: https://github.com/BaoBao0926/FTCFormer/tree/main.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

73 extracted references · 57 canonical work pages

  1. [1]

    Bian, F.-L

    Z. Bian, F.-L. Chung, and S. Wang. Fuzzy density peaks clustering. IEEE Transactions on Fuzzy Systems, 29(7):1725–1738, 2020

    work page 2020

  2. [2]

    Bolya, C.-Y

    D. Bolya, C.-Y . Fu, X. Dai, P. Zhang, C. Feichtenhofer, and J. Hoffman. Token merging: Your vit but faster. arXiv preprint arXiv:2210.09461, 2022

    arXiv 2022

  3. [3]

    Bossard, M

    L. Bossard, M. Guillaumin, and L. Van Gool. Food-101 – mining dis- criminative components with random forests. In European Conference on Computer Vision, 2014

    work page 2014

  4. [4]

    Carion, F

    N. Carion, F. Massa, G. Synnaeve, N. Usunier, A. Kirillov, and S. Zagoruyko. End-to-end object detection with transformers. In Euro- pean conference on computer vision, pages 213–229. Springer, 2020. Components Datasets DPC-FKNN SCS Cmerge Flowers102 DTD Stanford Cars Blood Cell Image GFLOPs Parameter (M) 77.83 52.66 81.83 88.73 3.84 14.23 ! 78.63 (+0.80)...

    work page 2020

  5. [5]

    W. Chen, C. Wang, Z. Zhang, Z. Huo, and L. Gao. Reweighted dynamic group convolution. In ICASSP 2021-2021 IEEE International Con- ference on Acoustics, Speech and Signal Processing (ICASSP) , pages 3940–3944. IEEE, 2021

    work page 2021

  6. [6]

    Cheng, J

    G. Cheng, J. Han, and X. Lu. Remote sensing image scene classifica- tion: Benchmark and state of the art. Proceedings of the IEEE, 105(10): 1865–1883, 2017

    work page 2017

  7. [7]

    Cimpoi, S

    M. Cimpoi, S. Maji, I. Kokkinos, S. Mohamed, , and A. Vedaldi. De- scribing textures in the wild. InProceedings of the IEEE Conf. on Com- puter Vision and Pattern Recognition (CVPR), 2014

    work page 2014

  8. [8]

    Coates, A

    A. Coates, A. Ng, and H. Lee. An analysis of single-layer networks in unsupervised feature learning. In Proceedings of the fourteenth inter- national conference on artificial intelligence and statistics, pages 215–

    work page

Show all 73 references
  1. [9]

    Dai and W

    D. Dai and W. Yang. Satellite image classification via two-layer sparse coding with biased image representation. IEEE Geoscience and remote sensing letters, 8(1):173–176, 2010

    2010

  2. [10]

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

    2009

  3. [12]

    S. Ding, W. Du, X. Xu, T. Shi, Y . Wang, and C. Li. An improved density peaks clustering algorithm based on natural neighbor with a merging strategy. Information Sciences, 624:252–276, 2023

    2023

  4. [13]

    Dosovitskiy, L

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

    2010 arXiv

  5. [14]

    M. Du, S. Ding, and H. Jia. Study on density peaks clustering based on k-nearest neighbors and principal component analysis. Knowledge- Based Systems, 99:135–145, 2016

    2016

  6. [15]

    M. Du, S. Ding, and Y . Xue. A robust density peaks clustering algorithm using fuzzy neighborhood. International Journal of Machine Learning and Cybernetics, 9:1131–1140, 2018

    2018

  7. [16]

    Fayyaz, S

    M. Fayyaz, S. A. Koohpayegani, F. R. Jafari, S. Sengupta, H. R. V . Joze, E. Sommerlade, H. Pirsiavash, and J. Gall. Adaptive token sampling for efficient vision transformers. In European Conference on Computer Vision, pages 396–414. Springer, 2022

    2022

  8. [17]

    Fei-Fei, R

    L. Fei-Fei, R. Fergus, and P. Perona. One-shot learning of object cate- gories. IEEE transactions on pattern analysis and machine intelligence, 28(4):594–611, 2006

    2006

  9. [18]

    I. J. Goodfellow, D. Erhan, P. L. Carrier, A. Courville, M. Mirza, B. Hamner, W. Cukierski, Y . Tang, D. Thaler, D.-H. Lee, et al. Chal- lenges in representation learning: A report on three machine learning contests. In Neural information processing: 20th international confer-...

    2013

  10. [19]

    Griffin, A

    G. Griffin, A. Holub, P. Perona, et al. Caltech-256 object category dataset. Technical report, Technical Report 7694, California Institute of Technology Pasadena, 2007

    2007

  11. [20]

    K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770–778, 2016

    2016

  12. [21]

    Helber, B

    P. Helber, B. Bischke, A. Dengel, and D. Borth. Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 12(7):2217–2226, 2019

    2019

  13. [22]

    A. G. Howard. Mobilenets: Efficient convolutional neural networks for mobile vision applications. arXiv preprint arXiv:1704.04861, 2017

    2017 arXiv

  14. [23]

    J. Howard. Imagenette: A smaller subset of imagenet for quick ex- periments, 2019. URL https://github.com/fastai/imagenette. Accessed: 2025-04-09

    2019

  15. [24]

    F. Ji, L. Li, T. Zhang, B. Zhang, J. Yang, J. Yin, and Q. Wang. A density peak clustering algorithm based on adaptive k-nearest neighbors with evidential strategy. InProceedings of the 2022 6th International Confer- ence on Computer Science and Artificial Intelligence , pages ...

    2022

  16. [25]

    M. Kim, S. Gao, Y .-C. Hsu, Y . Shen, and H. Jin. Token fusion: Bridging the gap between token pruning and token merging. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1383–1392, 2024

    2024

  17. [26]

    Z. Kong, P. Dong, X. Ma, X. Meng, W. Niu, M. Sun, X. Shen, G. Yuan, B. Ren, H. Tang, et al. Spvit: Enabling faster vision transformers via latency-aware soft token pruning. In European conference on computer vision, pages 620–640. Springer, 2022

    2022

  18. [27]

    Krause, M

    J. Krause, M. Stark, J. Deng, and L. Fei-Fei. 3d object representations for fine-grained categorization. In Proceedings of the IEEE interna- tional conference on computer vision workshops, pages 554–561, 2013

    2013

  19. [28]

    Krizhevsky, G

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

    2009

  20. [29]

    Le and X

    Y . Le and X. Yang. Tiny imagenet visual recognition challenge. CS 231N, 7(7):3, 2015

    2015

  21. [30]

    LeCun, L

    Y . LeCun, L. Bottou, Y . Bengio, and P. Haffner. Gradient-based learn- ing applied to document recognition. Proceedings of the IEEE, 86(11): 2278–2324, 1998

    1998

  22. [31]

    S. Lee, J. Choi, and H. J. Kim. Multi-criteria token fusion with one-step- ahead attention for efficient vision transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 15741–15750, 2024

    2024

  23. [32]

    J. Li, J. Liu, H. Yue, J. Cheng, H. Kuang, H. Bai, Y . Wang, and J. Wang. Darc: Deep adaptive regularized clustering for histopathological image classification. Medical image analysis, 80:102521, 2022

    2022

  24. [33]

    T. Li, P. Zhou, Z. He, X. Cheng, and X. Huang. Friendly sharpness- aware minimization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5631–5640, 2024

    2024

  25. [34]

    Y . Li, L. Sun, and Y . Tang. Dpc-fsc: An approach of fuzzy semantic cells to density peaks clustering. Information Sciences, 616:88–107, 2022

    2022

  26. [35]

    Y . Li, J. Hu, Y . Wen, G. Evangelidis, K. Salahi, Y . Wang, S. Tulyakov, and J. Ren. Rethinking vision transformers for mobilenet size and speed. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16889–16900, 2023

    2023

  27. [36]

    Liang, C

    Y . Liang, C. Ge, Z. Tong, Y . Song, J. Wang, and P. Xie. Not all patches are what you need: Expediting vision transformers via token reorgani- zations. arXiv preprint arXiv:2202.07800, 2022

    2022 arXiv

  28. [37]

    H. Liu, Z. Dai, D. So, and Q. V . Le. Pay attention to mlps. Advances in neural information processing systems, 34:9204–9215, 2021

    2021

  29. [38]

    Z. Liu, Y . Lin, Y . Cao, H. Hu, Y . Wei, Z. Zhang, S. Lin, and B. Guo. Swin transformer: Hierarchical vision transformer using shifted win- dows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012–10022, 2021

    2021

  30. [39]

    S. Long, Z. Zhao, J. Pi, S. Wang, and J. Wang. Beyond attentive tokens: Incorporating token importance and diversity for efficient vision trans- formers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10334–10343, 2023

    2023

  31. [40]

    Loshchilov and F

    I. Loshchilov and F. Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016

    2016 arXiv

  32. [41]

    Loshchilov and F

    I. Loshchilov and F. Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017

    2017 arXiv

  33. [42]

    Lotfi, P

    A. Lotfi, P. Moradi, and H. Beigy. Density peaks clustering based on density backbone and fuzzy neighborhood. Pattern Recognition, 107: 107449, 2020

    2020

  34. [43]

    Lu, S.-B

    W. Lu, S.-B. Chen, C. H. Ding, J. Tang, and B. Luo. Lwganet: A lightweight group attention backbone for remote sensing visual tasks. arXiv preprint arXiv:2501.10040, 2025

    2025

  35. [44]

    X. Ma, X. Dai, Y . Bai, Y . Wang, and Y . Fu. Rewrite the stars. In Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5694–5703, 2024

    2024

  36. [45]

    S. Maji, J. Kannala, E. Rahtu, M. Blaschko, and A. Vedaldi. Fine- grained visual classification of aircraft. Technical report, 2013

    2013

  37. [46]

    P. T. Mooney. Blood cell images, 2018. URL https://www.kaggle.com/ datasets/paultimothymooney/blood-cells. Accessed: 2025-04-09

    2018

  38. [47]

    Nilsback and A

    M.-E. Nilsback and A. Zisserman. Automated flower classification over a large number of classes. In 2008 Sixth Indian conference on computer vision, graphics & image processing, pages 722–729. IEEE, 2008

    2008

  39. [48]

    B. Pan, R. Panda, Y . Jiang, Z. Wang, R. Feris, and A. Oliva. Ia- red2: Interpretability-aware redundancy reduction for vision transform- ers. Advances in neural information processing systems , 34:24898– 24911, 2021

    2021

  40. [49]

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

    2012

  41. [50]

    Y . Rao, W. Zhao, B. Liu, J. Lu, J. Zhou, and C.-J. Hsieh. Dynam- icvit: Efficient vision transformers with dynamic token sparsification. Advances in neural information processing systems , 34:13937–13949, 2021

    2021

  42. [51]

    Rodriguez and A

    A. Rodriguez and A. Laio. Clustering by fast search and find of density peaks. science, 344(6191):1492–1496, 2014

    2014

  43. [52]

    X. Sun, J. Yang, M. Sun, and K. Wang. A benchmark for auto- matic visual classification of clinical skin disease images. In Com- puter Vision–ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part VI 14, pages 206–

    2016

  44. [53]

    Szegedy, W

    C. Szegedy, W. Liu, Y . Jia, P. Sermanet, S. Reed, D. Anguelov, D. Erhan, V . Vanhoucke, and A. Rabinovich. Going deeper with convolutions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1–9, 2015

    2015

  45. [54]

    Szegedy, V

    C. Szegedy, V . Vanhoucke, S. Ioffe, J. Shlens, and Z. Wojna. Rethink- ing the inception architecture for computer vision. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2818–2826, 2016

    2016

  46. [55]

    Vaswani, N

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, and I. Polosukhin. Attention is all you need. Ad- vances in neural information processing systems, 30, 2017

    2017

  47. [56]

    B. S. Veeling, J. Linmans, J. Winkens, T. Cohen, and M. Welling. Rota- tion equivariant cnns for digital pathology. InMedical image computing and computer assisted intervention–mICCAI 2018: 21st international conference, granada, Spain, September 16-20, 2018, proceedings, part...

    2018

  48. [57]

    H. Wang, S. Ge, Z. Lipton, and E. P. Xing. Learning robust global rep- resentations by penalizing local predictive power. Advances in neural information processing systems, 32, 2019

    2019

  49. [58]

    T. Wang, L. Yuan, Y . Chen, J. Feng, and S. Yan. Pnp-detr: Towards efficient visual analysis with transformers. In Proceedings of the IEEE/CVF international conference on computer vision , pages 4661– 4670, 2021

    2021

  50. [59]

    W. Wang, E. Xie, X. Li, D.-P. Fan, K. Song, D. Liang, T. Lu, P. Luo, and L. Shao. Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. InProceedings of the IEEE/CVF international conference on computer vision, pages 568–578, 2021

    2021

  51. [60]

    C. Wei, H. Fan, S. Xie, C.-Y . Wu, A. Yuille, and C. Feichtenhofer. Masked feature prediction for self-supervised visual pre-training. In Proceedings of the IEEE/CVF conference on computer vision and pat- tern recognition, pages 14668–14678, 2022

    2022

  52. [61]

    H. Wu, B. Xiao, N. Codella, M. Liu, X. Dai, L. Yuan, and L. Zhang. Cvt: Introducing convolutions to vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision , pages 22–31, 2021

    2021

  53. [62]

    E. Xie, W. Wang, Z. Yu, A. Anandkumar, J. M. Alvarez, and P. Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. Advances in neural information processing systems , 34: 12077–12090, 2021

    2021

  54. [63]

    S. Xie, R. Girshick, P. Dollár, Z. Tu, and K. He. Aggregated residual transformations for deep neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition , pages 1492– 1500, 2017

    2017

  55. [64]

    Y . Xu, Z. Zhang, M. Zhang, K. Sheng, K. Li, W. Dong, L. Zhang, C. Xu, and X. Sun. Evo-vit: Slow-fast token evolution for dynamic vision transformer. In Proceedings of the AAAI conference on artificial intelligence, volume 36, pages 2964–2972, 2022

    2022

  56. [65]

    Yang and S

    Y . Yang and S. Newsam. Bag-of-visual-words and spatial extensions for land-use classification. In Proceedings of the 18th SIGSPATIAL in- ternational conference on advances in geographic information systems, pages 270–279, 2010

    2010

  57. [66]

    H. Yin, A. Vahdat, J. M. Alvarez, A. Mallya, J. Kautz, and P. Molchanov. A-vit: Adaptive tokens for efficient vision transformer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10809–10818, 2022

    2022

  58. [67]

    S. Yun, D. Han, S. J. Oh, S. Chun, J. Choe, and Y . Yoo. Cutmix: Reg- ularization strategy to train strong classifiers with localizable features. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6023–6032, 2019

    2019

  59. [68]

    W. Zeng, S. Jin, W. Liu, C. Qian, P. Luo, W. Ouyang, and X. Wang. Not all tokens are equal: Human-centric visual analysis via token clustering transformer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11101–11111, 2022

    2022

  60. [69]

    W. Zeng, S. Jin, L. Xu, W. Liu, C. Qian, W. Ouyang, P. Luo, and X. Wang. Tcformer: Visual recognition via token clustering trans- former. IEEE Transactions on Pattern Analysis and Machine Intelli- gence, 2024

    2024

  61. [70]

    Zhang, M

    H. Zhang, M. Cisse, Y . N. Dauphin, and D. Lopez-Paz. mixup: Beyond empirical risk minimization. arXiv preprint arXiv:1710.09412, 2017

    2017 arXiv

  62. [71]

    J. Zhao, G. Wang, J.-S. Pan, T. Fan, and I. Lee. Density peaks cluster- ing algorithm based on fuzzy and weighted shared neighbor for uneven density datasets. Pattern Recognition, 139:109406, 2023

    2023

  63. [72]

    Zhao and Y

    X. Zhao and Y . Sun. Compress image to patches for vision transformer. arXiv preprint arXiv:2502.10120, 2025

    2025 arXiv

  64. [73]

    Zhong, L

    Z. Zhong, L. Zheng, G. Kang, S. Li, and Y . Yang. Random erasing data augmentation. In Proceedings of the AAAI conference on artificial intelligence, volume 34, pages 13001–13008, 2020

    2020

  65. [223]

    JMLR Workshop and Conference Proceedings, 2011

    2011

Pith tools

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