Conference paper · 2023 · Open access

MSViT: Dynamic Mixed-scale Tokenization for Vision Transformers

Jakob D. Havtorn, Amélie Royer, Tijmen Blankevoort, Babak Ehteshami Bejnordi

IEEE International Conference on Computer Vision Workshops, 838–848 · Published 2 October 2023 · DOI 10.1109/iccvw60793.2023.00091

Open PDFView at the publisherCite this paper

In brief

“MSViT: Dynamic Mixed-scale Tokenization for Vision Transformers” is a 2023 conference paper by Jakob D. Havtorn and 3 co-authors, published in IEEE International Conference on Computer Vision Workshops. OpenAlex counts 44 citations as of 22 November 2026, which puts it in the top 2% of works published in 2023 by citations. It is open access under a CC BY licence: the full text is free to read at arXiv (Cornell University).

Citations
44
Most cited in
2026 (27)
Field-weighted impact
2.56
References
58
Authors
4

Abstract

The input tokens to Vision Transformers carry little semantic meaning as they are defined as regular equal-sized patches of the input image, regardless of its content. However, processing uniform background areas of an image should not necessitate as much compute as dense, cluttered areas. To address this issue, we propose a dynamic mixed-scale tokenization scheme for ViT, MSViT. Our method introduces a conditional gating mechanism that selects the optimal token scale for every image region, such that the number of tokens is dynamically determined per input. In addition, to enhance the conditional behavior of the gate during training, we introduce a novel generalization of the batch-shaping loss. We show that our gating module is able to learn meaningful semantics despite operating locally at the coarse patch-level. The proposed gating module is lightweight, agnostic to the choice of transformer backbone, and trained within a few epochs with little training overhead. Furthermore, in contrast to token pruning, MSViT does not lose information about the input, thus can be readily applied for dense tasks. We validate MSViT on the tasks of classification and segmentation where it leads to improved accuracy-complexity trade-off.

Cite this paper

Formatted from the catalogue record. Check names, capitals and page numbers against the paper before you submit.

APA 7

Havtorn, J. D., Royer, A., Blankevoort, T., & Bejnordi, B. E. (2023). MSViT: Dynamic Mixed-scale Tokenization for Vision Transformers. IEEE International Conference on Computer Vision Workshops, 838–848. https://doi.org/10.1109/iccvw60793.2023.00091

MLA 9

Havtorn, Jakob D., et al. “MSViT: Dynamic Mixed-scale Tokenization for Vision Transformers.” IEEE International Conference on Computer Vision Workshops, 2023, pp. 838–848. https://doi.org/10.1109/iccvw60793.2023.00091.

Chicago

Havtorn, Jakob D., Amélie Royer, Tijmen Blankevoort, and Babak Ehteshami Bejnordi. 2023. “MSViT: Dynamic Mixed-scale Tokenization for Vision Transformers.” IEEE International Conference on Computer Vision Workshops: 838–848. https://doi.org/10.1109/iccvw60793.2023.00091.

Harvard

Havtorn, J. D. et al. (2023) ‘MSViT: Dynamic Mixed-scale Tokenization for Vision Transformers’, IEEE International Conference on Computer Vision Workshops, pp. 838–848. doi:10.1109/iccvw60793.2023.00091.

IEEE

J. D. Havtorn, A. Royer, T. Blankevoort, and B. E. Bejnordi, “MSViT: Dynamic Mixed-scale Tokenization for Vision Transformers,” IEEE International Conference on Computer Vision Workshops, pp. 838–848, 2023, doi: 10.1109/iccvw60793.2023.00091.

BibTeX

@inproceedings{havtorn2023msvit,
  title     = {{MSViT: Dynamic Mixed-scale Tokenization for Vision Transformers}},
  author    = {Havtorn, Jakob D. and Royer, Amélie and Blankevoort, Tijmen and Bejnordi, Babak Ehteshami},
  booktitle = {IEEE International Conference on Computer Vision Workshops},
  year      = {2023},
  pages     = {838--848},
  doi       = {10.1109/iccvw60793.2023.00091},
  url       = {https://doi.org/10.1109/iccvw60793.2023.00091}
}

Download for a reference manager: BibTeX (.bib) · RIS for Zotero, Mendeley and EndNote (.ris)

Citations by year

YearCitationsShare of the best year
202477
20251010
20262727

Its field-weighted citation impact is 2.56: it has been cited 2.56 times as often as the average work of the same field, type and year (1.00 is the world average).

Topic and keywords

Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Advanced Neural Network Applications

  • vision transformer
  • accuracy-complexity trade-off
  • image classification
  • semantic segmentation

Work with this paper

Most-cited papers in Computer Science

All highly cited Computer Science papers

Questions about this paper

How do I cite this paper in APA style?

Havtorn, J. D., Royer, A., Blankevoort, T., & Bejnordi, B. E. (2023). MSViT: Dynamic Mixed-scale Tokenization for Vision Transformers. IEEE International Conference on Computer Vision Workshops, 838–848. https://doi.org/10.1109/iccvw60793.2023.00091

How many times has this paper been cited?

44 times, according to OpenAlex on 22 November 2026. Google Scholar and Scopus count different sets of sources, so their numbers differ.

Is this paper free to read?

Yes. It is open access at arXiv (Cornell University): sign in and open the PDF from this page.

Who wrote this paper?

Jakob D. Havtorn, Amélie Royer, Tijmen Blankevoort, Babak Ehteshami Bejnordi.

What is the DOI of this paper?

10.1109/iccvw60793.2023.00091, permanently at https://doi.org/10.1109/iccvw60793.2023.00091. Sign in and open it from this page.

Bibliographic data from OpenAlex (CC0), last checked 22 November 2026. ResearchGuru does not host this paper; once you are signed in, the links open it at the publisher or the repository that does.