pith. sign in

arxiv: 2606.31676 · v1 · pith:YK5APSKSnew · submitted 2026-06-30 · 💻 cs.CV

REDI: Corpus Aware Patch Ranking for DINOv3 Token Reduction

Chanjong Im , Sebastian Diem , Thomas Mandl This is my paper

Pith reviewed 2026-07-01 05:55 UTC · model grok-4.3

classification 💻 cs.CV
keywords token reductionDINOv3patch rankingTF-IDFattention mapsImageNet classificationvision transformerefficiency
0
0 comments X

The pith

A class-conditioned corpus score combined with attention maps lets token reduction improve rather than degrade DINOv3 classification accuracy.

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

REDI addresses token reduction in Vision Transformers by asking how to best allocate a fixed budget of patch tokens given access to class labels during ranking. It builds a visual vocabulary from quantized DINOv3 features on validation data, computes supervised TF-IDF scores per class, and for each image multiplies the matching class row with an attention map to rank patches. The resulting ranking drives a keep-merge-compress operator that shrinks the sequence from 201 to 107 tokens. With a frozen backbone the approach reaches 84.706 percent top-1 accuracy on ImageNet-1K, exceeding the 83.514 percent of the full dense model. This matters because it demonstrates that corpus-level class statistics and per-image attention supply complementary information for deciding which patches to retain or merge.

Core claim

The paper establishes that the elementwise product of a class-conditioned TF-IDF map, obtained by quantizing final-block DINOv3 representations into a visual vocabulary and applying supervised TF-IDF over the corpus, and an attention map obtained from a separate dense forward pass, after independent min-max normalization, yields a patch ranking that, when used by a fixed keep-merge-compress operator, permits a 46.8 percent sequence-length reduction while increasing top-1 accuracy from 83.514 percent to 84.706 percent on ImageNet-1K relative to the dense baseline.

What carries the argument

The REDI score: the normalized product of a class-specific TF-IDF map derived from a supervised visual vocabulary and an attention map from the DINOv3 backbone.

If this is right

  • The combined signal outperforms using incoming attention mass alone (82.634 percent) or supervised TF-IDF alone (81.796 percent).
  • The same corpus term improves performance for three alternative attention formulations.
  • Precomputed REDI scores allow the reduction to be applied with a frozen backbone and unchanged linear classifier.
  • The operator assigns patch roles according to score rank and uses score magnitude to weight merging and compression.

Where Pith is reading between the lines

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

  • Rebuilding the visual vocabulary on a different dataset might allow the method to transfer to new domains without retraining the backbone.
  • The approach could be tested on whether the improvement persists when the linear classifier is also trained on the reduced sequences rather than kept fixed.
  • Extending the corpus scoring to object detection or segmentation tasks would require adapting the class conditioning to multiple labels per image.

Load-bearing premise

The assumption that class-conditioned TF-IDF scores from the validation corpus generalize to test images without overfitting or losing relevance when the backbone or dataset changes.

What would settle it

Observing that accuracy falls below the dense baseline when the REDI-ranked tokens are used on a held-out test set drawn from a different distribution would indicate the corpus scores do not generalize.

Figures

Figures reproduced from arXiv: 2606.31676 by Chanjong Im, Sebastian Diem, Thomas Mandl.

Figure 1
Figure 1. Figure 1: Overview of REDI. Four transformed views provide class-conditioned visual TF-IDF scores aligned to a reference center [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Visual word assignment distributions and REDI score mass. Left: visual word frequency by rank for separate [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
read the original abstract

Most token reduction methods for Vision Transformers seek favorable tradeoffs between accuracy and efficiency by pruning, merging, or pooling patch tokens. REDI (Relevance for DINOv3 Token Reduction) studies this question through a controlled supervised reference: how should a fixed token budget be allocated across patches for image classification? REDI quantizes final block DINOv3 patch representations into a visual vocabulary and derives class conditioned corpus scores using supervised TF-IDF over visual words. For each validation image, the ground truth class selects a row of the TF-IDF table, and four transformed views produce a TF-IDF map aligned to a reference center crop. A separate dense pass on the same crop provides an attention map. After independent min max normalization, their elementwise product defines the REDI score. A fixed keep, merge, and compress operator then uses score rank to assign patch roles and score magnitude to weight merging and compression. With precomputed REDI scores, a frozen DINOv3 ViT-B/16 backbone, and the same linear classifier used for dense evaluation, the operator reduces the sequence length from 201 to 107 tokens, a 46.8% sequence reduction. The REDI variant based on incoming attention mass achieves 84.706% Top-1 accuracy on ImageNet-1K, compared with 83.514% for the dense baseline, 82.634% for incoming attention mass alone, and 81.796% for supervised TF-IDF alone. The same corpus term also improves reduced classification for three alternative attention formulations relative to their attention only counterparts. Together, these controlled comparisons indicate that class specific corpus statistics and image specific attention provide complementary signals for patch ranking in this setting.

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 / 0 minor

Summary. The paper proposes REDI, a token reduction method for DINOv3 ViT-B/16 that quantizes patch tokens into a visual vocabulary, computes class-conditioned TF-IDF scores over the ImageNet-1K validation set, combines them (after min-max normalization) with an attention map from a separate dense pass, and uses the resulting scores to rank, merge, and compress tokens down to a fixed budget (e.g., 107 tokens from 201). With a frozen backbone and the same linear classifier, it reports 84.706% Top-1 accuracy versus 83.514% for the dense baseline, 82.634% for attention-only, and 81.796% for TF-IDF-only, claiming complementary signals from corpus statistics and attention.

Significance. If the numerical gains were obtained without label-dependent selection on the evaluation set, the result would indicate that class-specific corpus TF-IDF and image-specific attention provide complementary patch-ranking signals, offering a controlled reference point for token-reduction design in ViTs. The controlled ablations against TF-IDF-only and attention-only variants would strengthen that interpretation.

major comments (2)
  1. [Abstract] Abstract: the method explicitly states that 'for each validation image, the ground truth class selects a row of the TF-IDF table.' The TF-IDF table is constructed from the ImageNet-1K validation set, and all reported accuracies (including the 84.706% vs. 83.514% dense baseline comparison) are also measured on validation images. This supplies the token-ranking stage with ground-truth class labels unavailable at inference, so the observed improvement cannot be attributed solely to the REDI ranking operator.
  2. [Abstract] Abstract and evaluation description: no error bars, no statistical significance tests, and no details are provided on the choice of visual vocabulary size or the quantization procedure used to build the TF-IDF table. These omissions make it impossible to assess whether the 1.192-point gain over the dense baseline is robust or sensitive to those implementation choices.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the detailed comments. We address each major point below and indicate planned revisions to clarify the supervised nature of the method and to supply the requested implementation details.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the method explicitly states that 'for each validation image, the ground truth class selects a row of the TF-IDF table.' The TF-IDF table is constructed from the ImageNet-1K validation set, and all reported accuracies (including the 84.706% vs. 83.514% dense baseline comparison) are also measured on validation images. This supplies the token-ranking stage with ground-truth class labels unavailable at inference, so the observed improvement cannot be attributed solely to the REDI ranking operator.

    Authors: We agree that the formulation uses ground-truth class labels to index the TF-IDF table for each validation image. The manuscript explicitly frames REDI as a 'controlled supervised reference' whose purpose is to measure the best-case allocation of a fixed token budget when class identity is known. The reported gains therefore reflect the combination of corpus statistics and attention under this oracle condition rather than a label-free inference procedure. We will revise the abstract, introduction, and conclusion to state this limitation more explicitly, to avoid any implication that the 1.192-point improvement is achievable without class labels, and to discuss how the same corpus term might be adapted to an unsupervised or weakly supervised setting. revision: yes

  2. Referee: [Abstract] Abstract and evaluation description: no error bars, no statistical significance tests, and no details are provided on the choice of visual vocabulary size or the quantization procedure used to build the TF-IDF table. These omissions make it impossible to assess whether the 1.192-point gain over the dense baseline is robust or sensitive to those implementation choices.

    Authors: We will add the missing implementation details: the visual vocabulary was constructed with k-means (k = 8192) on a random subset of DINOv3 patch features from the ImageNet training set, followed by nearest-centroid assignment. We will also report results across three independent vocabulary constructions (different random seeds) and include standard deviations for the Top-1 accuracies. Because the backbone remains frozen, variance is modest, but the additional statistics will allow readers to judge robustness to vocabulary size and quantization. revision: yes

Circularity Check

0 steps flagged

No circularity in derivation chain

full rationale

The paper presents an empirical method that explicitly uses ground-truth class labels from the ImageNet-1K validation set to select rows from a precomputed class-conditioned TF-IDF table, then combines the result with an attention map for patch ranking. This is framed as a 'controlled supervised reference' for studying token allocation, with accuracies measured directly on the same set after applying the keep/merge/compress operator. No derivation steps, equations, or predictions reduce by construction to the inputs (no self-definitional relations, no fitted parameters renamed as predictions, and no self-citation load-bearing claims). The reported gains (e.g., 84.706% vs. 83.514% dense) are empirical outcomes from the described procedure rather than forced equivalences, and the method remains self-contained against its own benchmarks without invoking external uniqueness theorems or ansatzes.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The central claim rests on standard TF-IDF mathematics and the assumption that a fixed validation-derived corpus table remains useful at test time; no new physical entities are introduced and the only free parameter visible in the abstract is the target token count of 107.

free parameters (1)
  • target token count = 107
    The keep/merge/compress operator is applied to reach exactly 107 tokens; this budget is chosen for the reported experiment.
axioms (1)
  • domain assumption Quantized final-block DINOv3 features form a meaningful visual vocabulary for class discrimination via TF-IDF
    The method begins by quantizing patch representations into visual words and then applies supervised TF-IDF; this step is taken as given.

pith-pipeline@v0.9.1-grok · 5849 in / 1434 out tokens · 31594 ms · 2026-07-01T05:55:25.474002+00:00 · methodology

discussion (0)

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

Reference graph

Works this paper leans on

41 extracted references · 41 canonical work pages · 5 internal anchors

  1. [1]

    Quantifying attention flow in transformers

    Samira Abnar and Willem Zuidema. Quantifying attention flow in transformers. InProceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 4190–4197, 2020. 3, 6

    work page 2020

  2. [2]

    R\'enyi Attention Entropy for Patch Pruning

    Hiroaki Aizawa and Yuki Igaue. R ´enyi attention entropy for patch pruning, 2026. arXiv preprint arXiv:2604.03803. 3

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  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. InInternational Conference on Learning Representations, 2023. 1, 3, 6

    work page 2023

  4. [4]

    Structural Assessment for Understanding and Guiding Dataset Distillation in Discrete Token Space

    Yue Cao, Jianyang Gu, Vyacheslav Kungurtsev, Yu Hu, Jozsef Hamari, Zheng Liu, and Mohsen Zardadi. Struc- tural assessment for understanding and guiding dataset dis- tillation in discrete token space, 2026. arXiv preprint arXiv:2606.21705. 4

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  5. [5]

    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 IEEE/CVF International Conference on Computer Vision,

    work page

  6. [6]

    Dance, Lixin Fan, Jutta Willamowski, and C ´edric Bray

    Gabriella Csurka, Christopher R. Dance, Lixin Fan, Jutta Willamowski, and C ´edric Bray. Visual categorization with bags of keypoints. InECCV Workshop on Statistical Learn- ing in Computer Vision, pages 1–22, 2004. 1, 4

    work page 2004

  7. [7]

    Vision transformers need registers

    Timoth ´ee Darcet, Maxime Oquab, Julien Mairal, and Piotr Bojanowski. Vision transformers need registers. InInterna- tional Conference on Learning Representations, 2024. 2

    work page 2024

  8. [8]

    Dhillon and Dharmendra S

    Inderjit S. Dhillon and Dharmendra S. Modha. Concept de- compositions for large sparse text data using clustering.Ma- chine Learning, 42:143–175, 2001. 1, 4

    work page 2001

  9. [9]

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

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Syl- vain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. InInternational Conference on Learning Representa- tions, 2021. 1, 2, 4, 7

    work page 2021

  10. [10]

    Adaptive token sampling for efficient vision transformers

    Mohsen Fayyaz, Soroush Abbasi Koohpayegani, Farnoush Rezaei Jafari, Sunando Sengupta, Hamid Reza Vaezi Joze, Eric Sommerlade, Hamed Pirsiavash, and J¨urgen Gall. Adaptive token sampling for efficient vision transformers. InEuropean Conference on Computer Vision, pages 396–414, 2022. 3

    work page 2022

  11. [11]

    Visual-word tokenizer: Beyond fixed sets of to- kens in vision transformers.Transactions on Machine Learn- ing Research, 2025

    Leonidas Gee, Wing Yan Li, Viktoriia Sharmanska, and Novi Quadrianto. Visual-word tokenizer: Beyond fixed sets of to- kens in vision transformers.Transactions on Machine Learn- ing Research, 2025. 4

    work page 2025

  12. [12]

    Robustifying token attention for vision transformers

    Yong Guo, David Stutz, and Bernt Schiele. Robustifying token attention for vision transformers. InProceedings of the IEEE/CVF International Conference on Computer Vi- sion, pages 17557–17568, 2023. 1, 5

    work page 2023

  13. [13]

    Visual words meet BM25: Sparse auto-encoder visual word scoring for image retrieval, 2026

    Donghoon Han, Eunhwan Park, and Seunghyeon Seo. Visual words meet BM25: Sparse auto-encoder visual word scoring for image retrieval, 2026. 4

    work page 2026

  14. [14]

    Masked autoencoders are scalable vision learners

    Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Doll´ar, and Ross Girshick. Masked autoencoders are scalable vision learners. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16000– 16009, 2022. 8

    work page 2022

  15. [15]

    Rethinking Prototype-based Similarity Learning for Few-Shot Object Detection

    KunHo Heo, Seungjae Kim, Wongyu Lee, SuYeon Kim, and MyeongAh Cho. Rethinking prototype-based similarity learning for few-shot object detection, 2026. arXiv preprint arXiv:2606.23069. 4

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  16. [16]

    SPViT: Enabling faster vision transformers via latency-aware soft token prun- ing

    Zhenglun Kong, Peiyan Dong, Xiaolong Ma, Xin Meng, Wei Niu, Mengshu Sun, Xuan Shen, Geng Yuan, Bin Ren, Hao Tang, Minghai Qin, and Yanzhi Wang. SPViT: Enabling faster vision transformers via latency-aware soft token prun- ing. InEuropean Conference on Computer Vision, pages 620–640, 2022. 3

    work page 2022

  17. [17]

    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. InInternational Conference on Learning Representa- tions, 2022. 1, 3, 6

    work page 2022

  18. [18]

    Beyond attentive tokens: Incorporating to- ken importance and diversity for efficient vision transform- ers

    Sifan Long, Zhen Zhao, Jimin Pi, Shengsheng Wang, and Jingdong Wang. Beyond attentive tokens: Incorporating to- ken importance and diversity for efficient vision transform- ers. InIEEE/CVF Conference on Computer Vision and Pat- tern Recognition, pages 10334–10343, 2023. 3

    work page 2023

  19. [19]

    Token pooling in vision transformers for image classification

    Dmitrii Marin, Jen-Hao Rick Chang, Anurag Ranjan, An- ish Prabhu, Mohammad Rastegari, and Oncel Tuzel. Token pooling in vision transformers for image classification. In IEEE/CVF Winter Conference on Applications of Computer Vision, pages 12–21, 2023. 1, 3, 6

    work page 2023

  20. [20]

    AdaViT: Adaptive vision transformers for efficient image recognition

    Lingchen Meng, Hengduo Li, Bor-Chun Chen, Shiyi Lan, Zuxuan Wu, Yu-Gang Jiang, and Ser-Nam Lim. AdaViT: Adaptive vision transformers for efficient image recognition. InIEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12309–12318, 2022. 3

    work page 2022

  21. [21]

    Scalable recognition with a vocabulary tree

    David Nist ´er and Henrik Stew ´enius. Scalable recognition with a vocabulary tree. InIEEE Computer Society Con- ference on Computer Vision and Pattern Recognition, pages 2161–2168, 2006. 4

    work page 2006

  22. [22]

    DINOv2: Learning robust visual features without supervision.Transactions on Machine Learning Re- search, 2024

    Maxime Oquab, Timoth ´ee Darcet, Th ´eo Moutakanni, Huy V o, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Mah- moud Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Herv ´e J´egou, Julien Mairal, P...

    work page 2024

  23. [23]

    IA-RED 2: Interpretability-aware redundancy reduction for vision trans- formers

    Bowen Pan, Rameswar Panda, Yifan Jiang, Zhangyang Wang, Rogerio Feris, and Aude Oliva. IA-RED 2: Interpretability-aware redundancy reduction for vision trans- formers. InAdvances in Neural Information Processing Sys- tems, 2021. 3

    work page 2021

  24. [24]

    Object retrieval with large vocabularies and fast spatial matching

    James Philbin, Ondrej Chum, Michael Isard, Josef Sivic, and Andrew Zisserman. Object retrieval with large vocabularies and fast spatial matching. InIEEE Conference on Computer Vision and Pattern Recognition, pages 1–8, 2007. 4

    work page 2007

  25. [25]

    DynamicViT: Efficient vision transformers with dynamic token sparsification

    Yongming Rao, Wenliang Zhao, Benlin Liu, Jiwen Lu, Jie Zhou, and Cho-Jui Hsieh. DynamicViT: Efficient vision transformers with dynamic token sparsification. InAdvances in Neural Information Processing Systems, pages 13937– 13949, 2021. 1, 3

    work page 2021

  26. [26]

    Learning to merge tokens in vision transformers ,

    Cedric Renggli, Andr ´e Susano Pinto, Neil Houlsby, Basil Mustafa, Joan Puigcerver, and Carlos Riquelme. Learning to merge tokens in vision transformers, 2022. arXiv preprint arXiv:2202.12015. 3

    work page arXiv 2022

  27. [27]

    Understanding inverse document fre- quency: On theoretical arguments for IDF.Journal of Doc- umentation, 60(5):503–520, 2004

    Stephen Robertson. Understanding inverse document fre- quency: On theoretical arguments for IDF.Journal of Doc- umentation, 60(5):503–520, 2004. 1, 4, 5

    work page 2004

  28. [28]

    Berg, and Li Fei-Fei

    Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, San- jeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. ImageNet large scale visual recognition chal- lenge.International Journal of Computer Vision, 115(3): 211–252, 2015. 4, 7

    work page 2015

  29. [29]

    Ryoo, AJ Piergiovanni, Anurag Arnab, Mostafa Dehghani, and Anelia Angelova

    Michael S. Ryoo, AJ Piergiovanni, Anurag Arnab, Mostafa Dehghani, and Anelia Angelova. TokenLearner: Adaptive space-time tokenization for videos. InAdvances in Neural Information Processing Systems, 2021. 3

    work page 2021

  30. [30]

    Term-weighting ap- proaches in automatic text retrieval.Information Processing & Management, 24(5):513–523, 1988

    Gerard Salton and Christopher Buckley. Term-weighting ap- proaches in automatic text retrieval.Information Processing & Management, 24(5):513–523, 1988. 1, 4, 5

    work page 1988

  31. [31]

    Oriane Sim ´eoni, Huy V . V o, Maximilian Seitzer, Federico Baldassarre, Maxime Oquab, Cijo Jose, Vasil Khalidov, Marc Szafraniec, Seungeun Yi, Micha ¨el Ramamonjisoa, Francisco Massa, Daniel Haziza, Luca Wehrstedt, Jianyuan Wang, Timoth´ee Darcet, Th´eo Moutakanni, Leonel Sentana, Claire Roberts, Andrea Vedaldi, Jamie Tolan, John Brandt, Camille Couprie,...

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  32. [32]

    Video Google: A text retrieval approach to object matching in videos

    Josef Sivic and Andrew Zisserman. Video Google: A text retrieval approach to object matching in videos. InIEEE International Conference on Computer Vision, pages 1470– 1477, 2003. 1, 4, 5

    work page 2003

  33. [33]

    R ´enyi en- tropy: A new token pruning metric for vision transformers,

    Wei-Yuan Su, Ruijie Zhang, and Zheng Zhang. R ´enyi en- tropy: A new token pruning metric for vision transformers,

    work page

  34. [34]

    arXiv preprint arXiv:2603.27900. 3

    work page arXiv

  35. [35]

    Patch slimming for effi- cient vision transformers

    Yehui Tang, Kai Han, Yunhe Wang, Chang Xu, Jianyuan Guo, Chao Xu, and Dacheng Tao. Patch slimming for effi- cient vision transformers. InIEEE/CVF Conference on Com- puter Vision and Pattern Recognition, pages 12165–12174,

    work page

  36. [36]

    DeiT III: Revenge of the ViT

    Hugo Touvron, Matthieu Cord, and Herv ´e J´egou. DeiT III: Revenge of the ViT. InEuropean Conference on Computer Vision, pages 516–533. Springer, 2022. 8

    work page 2022

  37. [37]

    Teaching matters: Investigating the role of supervision in vision transformers

    Matthew Walmer, Saksham Suri, Kamal Gupta, and Abhi- nav Shrivastava. Teaching matters: Investigating the role of supervision in vision transformers. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7486–7496, 2023. 1, 6

    work page 2023

  38. [38]

    Federated unlearning via class-discriminative pruning

    Junxiao Wang, Song Guo, Xin Xie, and Heng Qi. Federated unlearning via class-discriminative pruning. InProceedings of the ACM Web Conference 2022, pages 622–632, 2022. 4

    work page 2022

  39. [39]

    Evo-ViT: Slow-fast token evolution for dynamic vision transformer

    Yifan Xu, Zhijie Zhang, Mengdan Zhang, Kekai Sheng, Ke Li, Weiming Dong, Liqing Zhang, Changsheng Xu, and Xing Sun. Evo-ViT: Slow-fast token evolution for dynamic vision transformer. InProceedings of the AAAI Conference on Artificial Intelligence, pages 2964–2972, 2022. 3

    work page 2022

  40. [40]

    Alvarez, Arun Mallya, Jan Kautz, and Pavlo Molchanov

    Hongxu Yin, Arash Vahdat, Jose M. Alvarez, Arun Mallya, Jan Kautz, and Pavlo Molchanov. A-ViT: Adaptive tokens for efficient vision transformer. InIEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10809– 10818, 2022. 3

    work page 2022

  41. [41]

    Faster or Stronger: Towards Flexible Visual Place Recognition via Weighted Aggregation and Token Pruning

    Zichao Zeng, June Moh Goo, Junwei Zheng, Weijia Fan, Ji- aming Zhang, Rainer Stiefelhagen, and Jan Boehm. Faster or stronger: Towards flexible visual place recognition via weighted aggregation and token pruning, 2026. arXiv preprint arXiv:2605.20551. 3

    work page internal anchor Pith review Pith/arXiv arXiv 2026