Pith. sign in

REVIEW 4 major objections 4 minor 65 references

A vision transformer for semantic segmentation can run entirely on integer arithmetic, staying within about 5% of full-precision accuracy.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-04 17:51 UTC pith:GHPKLYUN

load-bearing objection A solid engineering extension of I-ViT to segmentation—the lambda-ShiftGELU fix is real—but the 'fully integer-only' claim outruns what the paper actually demonstrates: PTQ is PyTorch-simulated and the TVM translation is not verified. the 4 major comments →

arxiv 2509.10334 v2 pith:GHPKLYUN submitted 2025-09-12 cs.CV cs.AIcs.LG

I-Segmenter: Integer-Only Vision Transformer for Efficient Semantic Segmentation

Jordan Sassoon , Michal Szczepanski , Martyna Poreba This is my paper
classification cs.CV cs.AIcs.LG
keywords semantic segmentationvision transformerinteger-only quantizationpost-training quantizationquantization-aware trainingGELU approximationedge deploymentlambda-ShiftGELU
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper tries to establish that a vision transformer for semantic segmentation can run entirely on integer arithmetic, with no floating-point operations anywhere in the graph, without collapsing in accuracy. Building on the Segmenter architecture, the authors replace every FP32 operator with an integer-only counterpart, remove L2 normalization, swap bilinear upsampling for nearest neighbor, and introduce a new GELU approximation called λ-ShiftGELU. The result is a model that stays within about 5% mIoU of its full-precision baseline on ADE20K and Cityscapes, shrinks the checkpoint by up to 3.8×, and can run up to 1.2× faster when deployed with a compiler that enforces integer kernels. The practical stake is edge deployment: integer-only models fit on low-power chips and avoid the need for floating-point hardware.

Core claim

I-Segmenter is the first fully integer-only vision transformer for semantic segmentation. Every operator—attention, feed-forward, normalization, softmax, GELU, and the decoder head—is quantized to integer arithmetic, with scale factors stored as dyadic fractions implemented by integer multiply and bit-shift. To prevent activation outliers from wrecking low-precision GELU, the paper proposes λ-ShiftGELU, which relaxes the lower clamp of the shift-based exponential approximation using a tunable parameter λ. Decoder-side, the L2 normalization is removed and bilinear interpolation is replaced with nearest-neighbor upsampling so the whole graph stays integer. In experiments, the model stays withi

What carries the argument

λ-ShiftGELU is the central new mechanism: an integer-only approximation of the GELU activation in which the exponential term is computed with bit shifts and the lower-bound clamp is relaxed by a tunable scalar λ. It replaces the overly restrictive clamp in the earlier ShiftGELU design and is what makes post-training quantization stable for larger backbones. Supporting machinery includes dyadic arithmetic (scale factors represented as integer numerators over power-of-two denominators, applied as multiply-then-shift), integer-only LayerNorm and Shiftmax softmax, INT16 residuals and attention accumulation, and two decoder simplifications—removal of L2 normalization and adoption of nearest-neigh

Load-bearing premise

The claim of true integer-only execution rests on a manually written translation from the original model code to a low-level compiler being complete and correct; the paper itself reports a 1–2 point mIoU drop from that translation, so an unnoticed floating-point fallback anywhere in the graph would invalidate the 'fully integer-only' claim.

What would settle it

Inspect the generated kernel assembly from I-Segmenter's compiled deployment and verify that every operation is integer: any floating-point load, store, or arithmetic instruction in the traced graph would falsify the central claim. Alternatively, rerun the one-shot PTQ experiment on a held-out larger model and check whether the mIoU gap to FP32 stays within the reported ~5%.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • ViT segmentation models can run on integer-only accelerators without a floating-point fallback, making edge deployment on low-power chips practical.
  • One-shot PTQ with a single calibration image completes in about one second, so a pretrained model can be adapted on-device with minimal calibration cost.
  • Checkpoint size drops by 3.2–3.8×, which eases storage and over-the-air updates for embedded systems.
  • The 1–2 point mIoU loss observed after kernel-level compilation indicates where future engineering effort should focus: custom integer kernels and more reliable model-to-compiler translation.
  • Accuracy stays within a moderate margin of the full-precision baseline, which broadens the range of segmentation applications that can tolerate integer-only deployment.

Where Pith is reading between the lines

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

  • Editorial inference: If λ-ShiftGELU's mechanism is the key stabilizer, the same relaxation idea could transfer to other long-tailed activations, such as SiLU/Swish, in integer-only vision transformers for classification, object detection, or depth estimation.
  • Editorial inference: The dependency on a hand-written compiler translation hints that mainstream inference runtimes will need native integer-only operator support before this approach reaches typical production pipelines; until then, the 'fully integer-only' claim is backend-specific.
  • Editorial inference: A natural stress test is to push below 8 bits; if the λ parameter can be tuned to compensate for 4-bit or 6-bit quantization, the framework might scale to ultra-low-precision segmentation, though the current experiments do not establish that.
  • Editorial inference: The ablation shows nearest-neighbor upsampling costs about 1.7 mIoU points on average; an integer-friendly upsampling method that preserves some boundary smoothness could close a large part of the remaining gap to the FP32 baseline.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 4 minor

Summary. The paper proposes I-Segmenter, an integer-only adaptation of the Segmenter vision transformer for semantic segmentation. Following the I-ViT quantization scheme, it replaces linear and nonlinear operators with integer-only counterparts, introduces a modified GELU approximation called λ-ShiftGELU, removes L2 normalization, and substitutes bilinear upsampling with nearest-neighbor interpolation. The authors evaluate PTQ and QAT variants on ADE20K and Cityscapes across four model sizes, report model-size and latency results across PyTorch, ONNX Runtime, TensorRT, and TVM, and claim that TVM is the only backend that executes the model with true integer-only kernels. The central claims are that I-Segmenter is the first fully integer-only ViT segmentation framework, that it stays within a reasonable mIoU margin of FP32, and that one-shot PTQ remains competitive.

Significance. If fully substantiated, this would be a useful contribution: it extends integer-only ViT quantization from classification to dense prediction, introduces an activation approximation that appears to stabilize PTQ, and provides a concrete deployment workflow with an honest comparison of backend behavior. The paper's strengths include the careful operator-level design, the ablation of decoder-side modifications, the demonstration of large model-size reductions (3.2–3.8×), and the measurement of memory traffic. The main value is practical: showing that an integer-only ViT segmentation pipeline is feasible on edge-oriented compilers such as TVM. However, several load-bearing claims currently rest on unverified or incomplete experimental support.

major comments (4)
  1. [§VI-A, Tables III/IV vs §VI-B, Table VII] The one-shot PTQ accuracy claim is not demonstrated under true integer-only execution. Section VI-A states that 'all mIoU scores obtained in PyTorch,' and Tables III/IV report PTQ-only results from the PyTorch simulation. Only the QAT model on ADE20K is evaluated in TVM (Table VII). Since the TVM translation itself incurs an accuracy drop (later discussed as 1–2 points), the PTQ results under real integer-only kernels are unknown and could exceed the claimed 5.1% average margin. Please provide TVM-based PTQ results, or explicitly scope the PTQ claim to simulated integer arithmetic.
  2. [§IV-B, Table VI] The 'fully integer-only via TVM' assertion is not verifiable from the paper. The only evidence is the checkmark in Table VI and a brief statement that TVM enforces integer-only execution 'down to the level of individual kernel computations.' No TVM graph dump, operator-to-kernel mapping, generated C/LLVM/TeD code, or fallback audit is provided. Because the PyTorch-to-TVM translation is manual and not released, and because ONNX Runtime and TensorRT are shown to fall back to FP32 silently, this claim needs concrete supporting artifacts or inspection results.
  3. [§III-B, Eq. (11), Table I] λ is a free hyperparameter set to 6, with k_inter set to 23, but no sensitivity analysis or selection procedure is given. Since λ-ShiftGELU is a central contribution and the dramatic PTQ improvements in Tables III/IV depend on it, the robustness of λ across datasets, model sizes, and bit-widths should be reported. A sweep over reasonable λ values (or a principled derivation) is needed to show that the improvement is not tuned to the evaluated configurations. Also, Eq. (11) uses k_inner while Algorithm 1 uses kinter; please unify notation.
  4. [§IV-A, Tables III/IV/VII] The experimental report lacks repeated runs, seeds, and error bars, and the provenance of the FP32 baselines is ambiguous. Section IV-A says checkpoints were 'either retrieved from the ALGM repository ... or generated by training the model from scratch,' without specifying which model sizes use which source. Given that some reported gaps are as small as 0.2–3.5 mIoU (e.g., Tiny QAT on ADE20K, TVM results in Table VII), variance and checkpoint provenance could affect the conclusions. Please report at least three runs with standard deviations and state the checkpoint source for each configuration.
minor comments (4)
  1. [Abstract and §I] The numerical claims are not fully consistent with the tables. The average one-shot PTQ drop across Tables III/IV is about 5.2 mIoU, not exactly 5.1, and the introduction's 'no more than 6.7% loss' is contradicted by the S=500 Cityscapes Base drop of 7.10 mIoU. Please reconcile.
  2. [§VI-B, Table VII] The text says the TVM translation causes a 'consistent 1-2 point drop' in mIoU, but comparing Table VII with the QAT rows of Table III gives drops of 0.21, 2.05, 0.84, and 0.75 for Tiny/Small/Base/Large. This is not a consistent 1–2 point drop; please describe the actual pattern.
  3. [Algorithm 1] The comment in Algorithm 1 refers to 'Eq. 10' for the lambda-adjusted clamp, but the equation is Eq. (11). Please correct the cross-reference.
  4. [§VI-B] Latency numbers in Table VI lack experimental methodology details: hardware, number of warm-up runs, repetitions, batch size, and input resolution. Since the paper emphasizes 1.2× speedups, these details are important for reproducibility.

Circularity Check

0 steps flagged

No significant circularity: the accuracy and efficiency results are empirical measurements against external baselines, and the few self-citations are not load-bearing.

full rationale

The paper's derivation chain is not circular. I-Segmenter is constructed by taking the public Segmenter architecture, adopting the external I-ViT quantization scheme as a stated starting point, modifying GELU via a tunable λ parameter, removing L2 normalization and bilinear upsampling, and then measuring mIoU on ADE20K and Cityscapes and latency/size across PyTorch, ONNX Runtime, TensorRT, and TVM. The λ parameter is a hyperparameter (set to 6 with k_inter=23), and its effect is evaluated by RMSE against the FP32 GELU and by downstream segmentation mIoU; no fitted parameter is renamed as a prediction. The paper's central accuracy claims are compared against an FP32 Segmenter baseline, not against a quantity that the method itself defines. The only co-author citation ([24], a pruning example in the related-work list) is not used to justify any load-bearing claim, and I-ViT is cited as external prior work rather than as the source of this paper's results. The assertion that TVM executes the model entirely with integer arithmetic is an engineering claim that is stated but not kernel-verified in the paper; that is a reproducibility or verification gap, not a circular step, because the claim does not reduce to the paper's own equations or fitted values.

Axiom & Free-Parameter Ledger

2 free parameters · 4 axioms · 0 invented entities

The method relies on standard quantization assumptions and one hand-picked hyperparameter pair (lambda, kinter). The most questionable assumption is the validity of the non-original baseline checkpoints, which directly affects the reported accuracy margins.

free parameters (2)
  • lambda = 6
    Tunable scalar that relaxes the lower-bound clamp in ShiftExp (Eq. 11). Set to 6 with no sensitivity analysis, likely tuned on validation data.
  • kinter = 23
    Intermediate bit-precision used in lambda-ShiftGELU. Chosen by hand along with lambda, no justification or ablation.
axioms (4)
  • standard math The dyadic approximation S ≈ b/2^c is accurate enough for INT8 inference.
    Equation 5 uses fixed-point arithmetic with right bit-shifts to approximate scaling factors; this is a standard approximation but its error is not quantified.
  • domain assumption EMA-based min/max clipping thresholds, frozen after warmup, remain valid for test data.
    Section IV-A describes estimating m via EMA and freezing it; any distribution shift at test time would produce out-of-range activations.
  • domain assumption Baseline Segmenter checkpoints from the ALGM repository or from-scratch training are representative of the original Segmenter.
    Section IV states the original releases are no longer available; the accuracy gap is measured against these non-original baselines.
  • ad hoc to paper lambda=6 is a valid global setting across all datasets, model sizes, and bit-widths.
    Equation 11 introduces lambda with no theoretical or empirical justification for the chosen value; only ShiftGELU vs lambda-ShiftGELU is compared, not different lambda values.

pith-pipeline@v1.3.0-alltime-deepseek · 19183 in / 10734 out tokens · 108128 ms · 2026-08-04T17:51:01.595435+00:00 · methodology

0 comments
read the original abstract

Vision Transformers (ViTs) have recently achieved strong results in semantic segmentation, yet their deployment on resource-constrained devices remains limited due to their high memory footprint and computational cost. Quantization offers an effective strategy to improve efficiency, but ViT-based segmentation models are notoriously fragile under low precision, as quantization errors accumulate across deep encoder-decoder pipelines. We introduce I-Segmenter, the first fully integer-only ViT segmentation framework. Building on the Segmenter architecture, I-Segmenter systematically replaces floating-point operations with integer-only counterparts. To further stabilize both training and inference, we propose $\lambda$-ShiftGELU, a novel activation function that mitigates the limitations of uniform quantization in handling long-tailed activation distributions. In addition, we remove the L2 normalization layer and replace bilinear interpolation in the decoder with nearest neighbor upsampling, ensuring integer-only execution throughout the computational graph. Extensive experiments show that I-Segmenter achieves accuracy within a reasonable margin of its FP32 baseline (5.1 % on average), while reducing model size by up to 3.8x and enabling up to 1.2x faster inference with optimized runtimes. Notably, even in one-shot PTQ with a single calibration image, I-Segmenter delivers competitive accuracy, underscoring its practicality for real-world deployment.

Figures

Figures reproduced from arXiv: 2509.10334 by Jordan Sassoon, Martyna Poreba, Michal Szczepanski.

Figure 1
Figure 1. Figure 1: Overview of our proposed model, I-Segmenter. The [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Visualization of activation patterns across different layer types in the fifth encoder block and first decoder block of [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Overview of the proposed I-MaskTransformer, fea [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Inference pipeline of I-Segmenter, from the PyTorch model to ONNX, TensorRT, and TVM backends, with FP32 and [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: GELU distribution comparison in the 10th encoder [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗

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

65 extracted references · 3 canonical work pages

  1. [1]

    Rethinking semantic segmentation from a sequence-to- sequence perspective with transformers,

    S. Zheng, J. Lu, H. Zhao, X. Zhu, Z. Luo, Y . Wang, Y . Fu, J. Feng, T. Xiang, P. H. Torr, and L. Zhang, “Rethinking semantic segmentation from a sequence-to- sequence perspective with transformers,” inCVPR, 2021. 1, 2

    2021

  2. [2]

    Vision transformers: From semantic segmentation to dense prediction,

    L. Zhang, J. Lu, S. Zheng, X. Zhao, X. Zhu, Y . Fu, T. Xiang, J. Feng, and P. H. S. Torr, “Vision transformers: From semantic segmentation to dense prediction,” International Journal of Computer Vision, vol. 132, no. 12, p. 6142–6162, Jul. 2024. [Online]. Available: http://dx.doi.org/10.1007/s11263-024-02173-w 2

    work page doi:10.1007/s11263-024-02173-w 2024

  3. [3]

    Vision trans- formers for dense prediction,

    R. Ranftl, A. Bochkovskiy, and V . Koltun, “Vision trans- formers for dense prediction,”ArXiv preprint, 2021. 2

    2021

  4. [4]

    Seg- menter: Transformer for semantic segmentation,

    R. Strudel, R. Garcia, I. Laptev, and C. Schmid, “Seg- menter: Transformer for semantic segmentation,” inPro- ceedings of the IEEE/CVF international conference on computer vision, 2021, pp. 7262–7272. 2

    2021

  5. [5]

    Segformer: Simple and efficient design for semantic segmentation with transformers,

    E. Xie, W. Wang, Z. Yu, A. Anandkumar, J. M. Alvarez, and P. Luo, “Segformer: Simple and efficient design for semantic segmentation with transformers,” 2021. [Online]. Available: https://arxiv.org/abs/2105.15203 2

    Pith/arXiv arXiv 2021

  6. [6]

    Segvit: Semantic segmentation with plain vision transformers,

    B. Zhang, Z. Tian, Q. Tang, X. Chu, X. Wei, C. Shen, and Y . Liu, “Segvit: Semantic segmentation with plain vision transformers,” 2022. [Online]. Available: https://arxiv.org/abs/2210.05844 3

    Pith/arXiv arXiv 2022

  7. [7]

    Segvitv2: Exploring efficient and continual semantic segmentation with plain vision transformers,

    B. Zhang, L. Liu, M. H. Phan, Z. Tian, C. Shen, and Y . Liu, “Segvitv2: Exploring efficient and continual semantic segmentation with plain vision transformers,”

  8. [8]

    Structtoken: Rethinking semantic segmentation with structural prior,

    F. Lin, Z. Liang, S. Wu, J. He, K. Chen, and S. Tian, “Structtoken: Rethinking semantic segmentation with structural prior,”IEEE Transactions on Circuits and Systems for Video Technology, vol. 33, no. 10, pp. 5655– 5663, 2023. 2

    2023

  9. [9]

    Per- pixel classification is not all you need for semantic segmentation,

    B. Cheng, A. G. Schwing, and A. Kirillov, “Per- pixel classification is not all you need for semantic segmentation,” 2021. [Online]. Available: https://arxiv. org/abs/2107.06278 3

    Pith/arXiv arXiv 2021

  10. [10]

    Masked-attention mask transformer for uni- versal image segmentation,

    B. Cheng, I. Misra, A. G. Schwing, A. Kirillov, and R. Girdhar, “Masked-attention mask transformer for uni- versal image segmentation,”arXiv, 2021. 3

    2021

  11. [11]

    Your vit is secretly an image segmentation model,

    T. Kerssies, N. Cavagnero, A. Hermans, N. Norouzi, G. Averta, B. Leibe, G. Dubbelman, and D. de Geus, “Your vit is secretly an image segmentation model,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025. 1, 3

    2025

  12. [12]

    Attention is all you need,

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, and I. Polosukhin, “Attention is all you need,”Advances in neural informa- tion processing systems, vol. 30, 2017. 1

    2017

  13. [13]

    Understanding and overcoming the challenges of efficient transformer quantization,

    Y . Bondarenko, M. Nagel, and T. Blankevoort, “Understanding and overcoming the challenges of efficient transformer quantization,” 2021. [Online]. Available: https://arxiv.org/abs/2109.12948 1

    Pith/arXiv arXiv 2021

  14. [14]

    A survey of techniques for optimizing transformer inference,

    K. T. Chitty-Venkata, S. Mittal, M. Emani, V . Vishwanath, and A. K. Somani, “A survey of techniques for optimizing transformer inference,” 2023. [Online]. Available: https://arxiv.org/abs/2307.07982

    Pith/arXiv arXiv 2023

  15. [15]

    A comprehensive survey on recent model compression and acceleration approaches for deep neural networks and transformers,

    G. XU, W. Huang, and W. Jia, “A comprehensive survey on recent model compression and acceleration approaches for deep neural networks and transformers,” Jul. 2024, available at SSRN: https://ssrn.com/abstract= 4893335 or http://dx.doi.org/10.2139/ssrn.4893335

    work page doi:10.2139/ssrn.4893335 2024

  16. [16]

    Vision transformers on the edge: A comprehensive survey of model compression and acceler- ation strategies,

    S. Saha and L. Xu, “Vision transformers on the edge: A comprehensive survey of model compression and acceler- ation strategies,”Neurocomputing, vol. 643, p. 130417,

  17. [17]

    Mobilevit: Light-weight, general-purpose, and mobile-friendly vision transformer,

    S. Mehta and M. Rastegari, “Mobilevit: Light-weight, general-purpose, and mobile-friendly vision transformer,”

  18. [18]

    Rethinking vision transformers for mobilenet size and speed,

    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,” 2023. [Online]. Available: https://arxiv.org/abs/2212.08059

    Pith/arXiv arXiv 2023

  19. [19]

    Edgevits: Competing light-weight cnns on mobile devices with vision transformers,

    J. Pan, A. Bulat, F. Tan, X. Zhu, L. Dudziak, H. Li, G. Tzimiropoulos, and B. Martinez, “Edgevits: Competing light-weight cnns on mobile devices with vision transformers,” 2022. [Online]. Available: https: //arxiv.org/abs/2205.03436

    Pith/arXiv arXiv 2022

  20. [20]

    Mobile-former: Bridging mobilenet and transformer,

    Y . Chen, X. Dai, D. Chen, M. Liu, X. Dong, L. Yuan, and Z. Liu, “Mobile-former: Bridging mobilenet and transformer,” 2022. [Online]. Available: https://arxiv.org/abs/2108.05895

    Pith/arXiv arXiv 2022

  21. [21]

    Microvit: A vision transformer with low complexity self attention for edge device,

    N. Setyawan, C.-C. Sun, M.-H. Hsu, W.-K. Kuo, and J.-W. Hsieh, “Microvit: A vision transformer with low complexity self attention for edge device,” 02 2025. 1

    2025

  22. [22]

    A lightweight vision transformer with weighted global average pooling: implications for iomt applications,

    H. Dong, I. Kotenko, and S. Dong, “A lightweight vision transformer with weighted global average pooling: implications for iomt applications,”Complex & Intelligent Systems, vol. 11, p. 215, 2025. [Online]. Available: https://doi.org/10.1007/s40747-025-01842-8

    work page doi:10.1007/s40747-025-01842-8 2025

  23. [23]

    Algm: Adaptive local-then-global token merging for efficient semantic segmentation with plain vision trans- formers,

    N. Norouzi, S. Orlova, D. De Geus, and G. Dubbelman, 12 “Algm: Adaptive local-then-global token merging for efficient semantic segmentation with plain vision trans- formers,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 15 773–15 782. 1, 7

    2024

  24. [24]

    Step: Supertoken and early-pruning for efficient semantic segmentation,

    M. Proust, M. Poreba, M. Szczepanski, and K. Haroun, “Step: Supertoken and early-pruning for efficient semantic segmentation,” inVISIGRAPP 2025-20th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications, 2025, pp. 56–61. [Online]. Available: https: //www.scitepress.org/Papers/2025/131328/131328.pdf

    2025

  25. [25]

    Prune and merge: Efficient token compression for vision transformer with spatial information preserved,

    J. Mao, Y . Shen, J. Guo, Y . Yao, X. Hua, and H. Shen, “Prune and merge: Efficient token compression for vision transformer with spatial information preserved,”IEEE Transactions on Multimedia, vol. PP, pp. 1–14, 01 2025

    2025

  26. [26]

    Ucc: A unified cascade compression framework for vision transformer models,

    D. Chen, K. Lin, and Q. Deng, “Ucc: A unified cascade compression framework for vision transformer models,” Neurocomputing, vol. 612, p. 128747, 2025. [Online]. Available: https://www.sciencedirect.com/science/article/ pii/S0925231224015182

    2025

  27. [27]

    Atom: Adaptive token merging for efficient acceleration of vision transformer,

    J. Shin, M. Kang, Y . Han, J. Park, and L.-S. Kim, “Atom: Adaptive token merging for efficient acceleration of vision transformer,”IEEE Transactions on Computers, vol. 74, no. 5, pp. 1620–1633, 2025

    2025

  28. [28]

    Efficient token pruning in vision transformers using an attention-based multilayer network,

    M. Marchetti, D. Traini, D. Ursino, and L. Virgili, “Efficient token pruning in vision transformers using an attention-based multilayer network,”Expert Systems with Applications, vol. 279, p. 127449, 2025. [Online]. Available: https://www.sciencedirect.com/science/article/ pii/S0957417425010711

    2025

  29. [29]

    Revisiting token pruning for object detection and instance segmentation,

    Y . Liu, M. Gehrig, N. Messikommer, M. Cannici, and D. Scaramuzza, “Revisiting token pruning for object detection and instance segmentation,” in2024 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). Los Alamitos, CA, USA: IEEE Computer Society, Jan. 2024, pp. 2646–2656. [Online]. Available: https://doi.ieeecomputersociety.org/ 10.1109/...

    arXiv 2024

  30. [30]

    Adalog: Post-training quantization for vision transform- ers with adaptive logarithm quantizer,

    Z. Wu, J. Chen, H. Zhong, D. Huang, and Y . Wang, “Adalog: Post-training quantization for vision transform- ers with adaptive logarithm quantizer,” inEuropean Con- ference on Computer Vision (ECCV), 2024. 1, 3

    2024

  31. [31]

    Easyquant: An efficient data-free quantization algorithm for llms,

    H. Tang, Y . Sun, D. Wu, K. Liu, J. Zhu, and Z. Kang, “Easyquant: An efficient data-free quantization algorithm for llms,” 2024. [Online]. Available: https: //arxiv.org/abs/2403.02775

    Pith/arXiv arXiv 2024

  32. [32]

    Smoothquant: Accurate and efficient post- training quantization for large language models,

    G. Xiao, J. Lin, M. Seznec, H. Wu, J. Demouth, and S. Han, “Smoothquant: Accurate and efficient post- training quantization for large language models,” 2024. [Online]. Available: https://arxiv.org/abs/2211.10438

    Pith/arXiv arXiv 2024

  33. [33]

    Ptq4vit: Post-training quantization for vision transformers with twin uniform quantization,

    Z. Yuan, C. Xue, Y . Chen, Q. Wu, and G. Sun, “Ptq4vit: Post-training quantization for vision transformers with twin uniform quantization,” 2024. [Online]. Available: https://arxiv.org/abs/2111.12293 3, 4

    Pith/arXiv arXiv 2024

  34. [34]

    Mixed non-linear quantization for vision transformers,

    G. Kim, J. Lee, S. Park, Y . Kwon, and H. Kim, “Mixed non-linear quantization for vision transformers,” 2SETR025. [Online]. Available: https://arxiv.org/abs/ 2407.18437

    Pith/arXiv arXiv

  35. [35]

    Fima- q: Post-training quantization for vision transformers by fisher information matrix approximation,

    Z. Wu, S. Wang, J. Zhang, J. Chen, and Y . Wang, “Fima- q: Post-training quantization for vision transformers by fisher information matrix approximation,” inIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025. 2, 3, 4

    2025

  36. [36]

    Orq-vit: Outlier resilient post training quantization for vision transformers via outlier decomposition,

    X. He, Y . Lu, H. Liu, C. Gong, and W. He, “Orq-vit: Outlier resilient post training quantization for vision transformers via outlier decomposition,” Journal of Systems Architecture, vol. 168, p. 103530,

  37. [37]

    Training data-efficient image transformers & distillation through attention,

    H. Touvron, M. Cord, M. Douze, F. Massa, A. Sablay- rolles, and H. J ´egou, “Training data-efficient image transformers & distillation through attention,” 2021. [Online]. Available: https://arxiv.org/abs/2012.12877 1

    Pith/arXiv arXiv 2021

  38. [38]

    Knowledge distillation in vision transformers: A critical review,

    G. Habib, T. J. Saleem, and B. Lall, “Knowledge distillation in vision transformers: A critical review,”

  39. [39]

    Imagenet: A large-scale hierarchical image database,

    J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei, “Imagenet: A large-scale hierarchical image database,” in2009 IEEE Conference on Computer Vision and Pattern Recognition, 2009, pp. 248–255. 2

    2009

  40. [40]

    Available: https://www.sciencedirect

    [Online]. Available: https://www.sciencedirect. com/science/article/pii/S1383762125002024 1, 3, 4

  41. [41]

    Nvidia tensorrt,

    N. T. developers, “Nvidia tensorrt,” https://developer. nvidia.com/tensorrt, 2021, version: x.y.z. 2

    2021

  42. [42]

    Tvm: An automated end-to-end optimizing compiler for deep learning,

    T. Chen, T. Moreau, Z. Jiang, L. Zheng, E. Yan, M. Cowan, H. Shen, L. Wang, Y . Hu, L. Ceze, C. Guestrin, and A. Krishnamurthy, “Tvm: An automated end-to-end optimizing compiler for deep learning,” 2018. [Online]. Available: https://arxiv.org/abs/1802.04799 2

    Pith/arXiv arXiv 2018

  43. [43]

    Onnx runtime,

    O. R. developers, “Onnx runtime,” https: //www.onnxruntime.ai, 2021, version: x.y.z. 2

    2021

  44. [44]

    Fq- vit: Post-training quantization for fully quantized vision transformer,

    Y . Lin, T. Zhang, P. Sun, Z. Li, and S. Zhou, “Fq- vit: Post-training quantization for fully quantized vision transformer,” inProceedings of the Thirty-First Interna- tional Joint Conference on Artificial Intelligence, IJCAI- 22, 2022, pp. 1173–1179. 3

    2022

  45. [45]

    I-vit: Integer-only quantization for efficient vision transformer inference,

    Z. Li and Q. Gu, “I-vit: Integer-only quantization for efficient vision transformer inference,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 17 065–17 075. 2, 3

    2023

  46. [46]

    Repq-vit: Scale reparameterization for post-training quantization of vision transformers,

    Z. Li, J. Xiao, L. Yang, and Q. Gu, “Repq-vit: Scale reparameterization for post-training quantization of vision transformers,” 2023. [Online]. Available: https://arxiv.org/abs/2212.08254 4

    Pith/arXiv arXiv 2023

  47. [47]

    Aphq-vit: Post-training quantization with aver- age perturbation hessian based reconstruction for vision transformers,

    Z. Wu, J. Zhang, J. Chen, J. Guo, D. Huang, and Y . Wang, “Aphq-vit: Post-training quantization with aver- age perturbation hessian based reconstruction for vision transformers,” inIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025. 3, 4 13

    2025

  48. [48]

    Adfq-vit: Activation-distribution-friendly post-training quantization for vision transformers,

    Y . Jiang, N. Sun, X. Xie, F. Yang, and T. Li, “Adfq-vit: Activation-distribution-friendly post-training quantization for vision transformers,” 2024. [Online]. Available: https://arxiv.org/abs/2407.02763

    Pith/arXiv arXiv 2024

  49. [49]

    Towards Accurate Post-Training Quantization of Vision Transformers via Error Reduction ,

    Y . Zhong, Y . Huang, J. Hu, Y . Zhang, and R. Ji, “ Towards Accurate Post-Training Quantization of Vision Transformers via Error Reduction ,” IEEE Transactions on Pattern Analysis & Machine Intelligence, vol. 47, no. 04, pp. 2676–2692, Apr. 2025. [Online]. Available: https://doi.ieeecomputersociety.org/ 10.1109/TPAMI.2025.3528042 3

    arXiv 2025

  50. [50]

    Towards accurate post-training quantization for vision transformer,

    Y . Ding, H. Qin, Q. Yan, Z. Chai, J. Liu, X. Wei, and X. Liu, “Towards accurate post-training quantization for vision transformer,” inProceedings of the 30th ACM International Conference on Multimedia, ser. MM ’22. New York, NY , USA: Association for Computing Machinery, 2022, p. 5380–5388. [Online]. Available: https://doi.org/10.1145/3503161.3547826 3

    arXiv 2022

  51. [51]

    Bivit: Extremely compressed binary vision transformers,

    Y . He, Z. Lou, L. Zhang, J. Liu, W. Wu, H. Zhou, and B. Zhuang, “Bivit: Extremely compressed binary vision transformers,” inProceedings - 2023 IEEE/CVF International Conference on Computer Vision, ICCV 2023, F. Jurie and G. Sharma, Eds. United States of America: IEEE, Institute of Electrical and Electronics Engineers, 2023, pp. 5628–5640, funding Informa...

    arXiv 2023

  52. [52]

    Packqvit: faster sub-8-bit vision transformers via full and packed quantization on the mobile,

    P. Dong, L. Lu, C. Wu, C. Lyu, G. Yuan, H. Tang, and Y . Wang, “Packqvit: faster sub-8-bit vision transformers via full and packed quantization on the mobile,” in Proceedings of the 37th International Conference on Neural Information Processing Systems, ser. NIPS ’23. Red Hook, NY , USA: Curran Associates Inc., 2023. 3, 4

    2023

  53. [53]

    Mix-qvit: Mixed- precision vision transformer quantization driven by layer importance and quantization sensitivity,

    N. Ranjan and A. Savakis, “Mix-qvit: Mixed- precision vision transformer quantization driven by layer importance and quantization sensitivity,” 2025. [Online]. Available: https://arxiv.org/abs/2501.06357 3

    Pith/arXiv arXiv 2025

  54. [54]

    Model quantization and hardware acceleration for vision transformers: A comprehensive survey,

    D. Du, G. Gong, and X. Chu, “Model quantization and hardware acceleration for vision transformers: A comprehensive survey,” 2024. 3

    2024

  55. [55]

    Q-vit: accurate and fully quantized low-bit vision trans- former,

    Y . Li, S. Xu, B. Zhang, X. Cao, P. Gao, and G. Guo, “Q-vit: accurate and fully quantized low-bit vision trans- former,” inProceedings of the 36th International Con- ference on Neural Information Processing Systems, ser. NIPS ’22. Red Hook, NY , USA: Curran Associates Inc.,

  56. [56]

    Quantiza- tion and training of neural networks for efficient integer- arithmetic-only inference,

    B. Jacob, S. Kligys, B. Chen, M. Zhu, M. Tang, A. Howard, H. Adam, and D. Kalenichenko, “Quantiza- tion and training of neural networks for efficient integer- arithmetic-only inference,” inProceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2018. 5

    2018

  57. [57]

    Scene parsing through ade20k dataset,

    B. Zhou, H. Zhao, X. Puig, S. Fidler, A. Barriuso, and A. Torralba, “Scene parsing through ade20k dataset,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2017. 7

    2017

  58. [58]

    The cityscapes dataset for semantic urban scene under- standing,

    M. Cordts, M. Omran, S. Ramos, T. Rehfeld, M. En- zweiler, R. Benenson, U. Franke, S. Roth, and B. Schiele, “The cityscapes dataset for semantic urban scene under- standing,” inProc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 7

    2016

  59. [59]

    Cyclical learning rates for training neural networks,

    L. N. Smith, “Cyclical learning rates for training neural networks,” in2017 IEEE winter conference on applica- tions of computer vision (WACV). IEEE, 2017, pp. 464–

    2017

  60. [60]

    End-to-end object detection with transformers,

    N. Carion, F. Massa, G. Synnaeve, N. Usunier, A. Kir- illov, and S. Zagoruyko, “End-to-end object detection with transformers,” inComputer Vision – ECCV 2020, A. Vedaldi, H. Bischof, T. Brox, and J.-M. Frahm, Eds. Cham: Springer International Publishing, 2020, pp. 213–

    2020

  61. [65]

    MMSegmentation: Openmmlab se- mantic segmentation toolbox and benchmark,

    M. Contributors, “MMSegmentation: Openmmlab se- mantic segmentation toolbox and benchmark,” https:// github.com/open-mmlab/mmsegmentation, 2020. 7 Jordan Sassoonis a Research Engineer at CEA-List and a MSc candidate in Data and Artificial Intelli- gence at Institut Polytechnique de Paris, specializing in Computer Vision. He received his BSc at Technis- ch...

    2020

  62. [2022]

    Available: https://arxiv.org/abs/2110

    [Online]. Available: https://arxiv.org/abs/2110. 02178 1

  63. [2023]

    Available: https://arxiv.org/abs/2306

    [Online]. Available: https://arxiv.org/abs/2306. 06289 3

  64. [2024]

    Available: https://arxiv.org/abs/2302

    [Online]. Available: https://arxiv.org/abs/2302. 02108 1

  65. [2025]

    Available: https://www.sciencedirect

    [Online]. Available: https://www.sciencedirect. com/science/article/pii/S0925231225010896 1