Pith. sign in

REVIEW 3 major objections 5 minor 54 references

A hybrid CNN-Mamba backbone claims 82.6% top-1 on ImageNet-1K with 31.5 million parameters.

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 ↗

VCMamba reports that using convolutional feed-forward blocks for the first three stages followed by multi-directional Mamba blocks in the final stage yields 82.6% ImageNet-1K and 47.1 ADE20K mIoU at 31.5M parameters, beating PlainMamba-L3, ViG-B, and EfficientFormer-L7 at fewer parameters.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection An honest CNN-Mamba hybrid with real ablations, but the headline margins rest on un-re-run baselines; should be reviewed, not desk-rejected. the 3 major comments →

arxiv 2509.04669 v1 pith:FTVN7W7C submitted 2025-09-04 cs.CV cs.AIcs.LG

VCMamba: Bridging Convolutions with Multi-Directional Mamba for Efficient Visual Representation

Mustafa Munir , Alex Zhang , Radu Marculescu This is my paper
classification cs.CV cs.AIcs.LG
keywords vision backbonestate space modelsMambaCNN-SSM hybridImageNet-1KADE20K semantic segmentationmulti-directional scanninglinear complexity
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 reading

The paper tries to establish that a hybrid architecture—convolutional stem and early stages, multi-directional Mamba stage at the end—outperforms pure CNNs, pure vision SSMs, ViTs, and graph-based backbones on accuracy per parameter. It reports VCMamba-B at 82.6% top-1 on ImageNet-1K with 31.5M parameters, beating PlainMamba-L3 by 0.3 points with 37% fewer parameters, and 47.1 mIoU on ADE20K, beating EfficientFormer-L7 by 2.0 points with 62% fewer backbone parameters. The sympathetic reading is that local feature extraction should happen at high resolution where convolutions are cheap and detail-rich, and global modeling should happen only after downsampling, where Mamba's linear-time selective scan can still propagate context across the whole feature map. If true, this gives a design template for efficient backbones that scale to high-resolution inputs without the quadratic cost of self-attention.

Core claim

VCMamba claims a better accuracy-per-parameter trade-off when convolutional blocks take the early high-resolution stages and a multi-directional Mamba stage takes the final low-resolution stage. Concretely: VCMamba-B hits 82.6% top-1 on ImageNet-1K with 31.5M parameters, 0.3 points above PlainMamba-L3 with 37% fewer parameters and 0.3 above ViG-B with 64% fewer; on ADE20K with Semantic FPN it hits 47.1 mIoU, 2.0 above EfficientFormer-L7 with 62% fewer backbone parameters. The pipeline is a two-convolution stem, three stages of inverted-residual-style FFN blocks, then a final stage interleaving FFNs with four-direction continuous-scan Mamba blocks with direction-aware updates. The story is th

What carries the argument

The Multi-Directional Mamba block (the paper's named module) plus the hierarchical convolutional stem. The block adapts Mamba's selective scan to 2D by flattening tokens along four snake-like paths (row and column), injecting learnable per-direction parameters into the SSM's input-dependent B update, summing the four outputs, then normalizing and projecting with convolutions. Its job is to give the final stage a global receptive field without quadratic attention, operating on already-rich local features produced by the early conv stages.

Load-bearing premise

The headline margins depend on the published PlainMamba-L3, ViG-B, and EfficientFormer-L7 numbers being comparable to VCMamba's own 300-epoch training recipe; the paper retrains no external baseline under its own pipeline.

What would settle it

Retrain PlainMamba-L3, ViG-B, and EfficientFormer-L7 with VCMamba's exact ImageNet schedule and ADE20K Semantic FPN fine-tuning, then compare. If a single directly re-run baseline matches or exceeds VCMamba-B on both tasks, the stated efficiency and accuracy margins fail. A second check: replace VCMamba-B's four Mamba blocks with equal-cost conv or attention blocks; if accuracy does not drop, the Mamba mechanism is not carrying the claimed gain.

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

If this is right

  • If the reported numbers hold, VCMamba-B improves accuracy per parameter over both a non-hierarchical Mamba model (PlainMamba-L3) and a vision GNN (ViG-B) on ImageNet-1K.
  • The same backbone transfers to dense prediction: with Semantic FPN, VCMamba-B reaches 47.1 mIoU on ADE20K, 2.0 points above the much larger EfficientFormer-L7.
  • Because the Mamba stage runs after heavy downsampling and uses linear-time scans, the architecture can be applied to higher-resolution inputs without self-attention's quadratic blow-up.
  • The paper's incremental ablations attribute a +2.4% top-1 gain over its CNN-Mamba baseline to the VCMamba block design: skip connection, interleaved Mamba-FFN, LayerNorm, convolutional projections, and stage-wise batch normalization.

Where Pith is reading between the lines

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

  • A matched-protocol rerun of the classification and segmentation baselines is the natural stress test; without it, the +0.3% and +2.0 mIoU margins rest on published numbers produced under different training schedules, augmentations, and decoders.
  • The architecture implies a scheduling principle—spend cheap local operators on high-resolution stages and apply global sequence modeling only after heavy downsampling—that could transfer to other SSM families or even to other global operators.
  • One could isolate whether Mamba is load-bearing by replacing the final MDM blocks with equal-parameter conv or attention blocks; the paper's ablations add components to a CNN-Mamba baseline but never remove Mamba.
  • The strided convolutional stem replaces patch embedding, so part of the gain may come from avoiding patchification information loss; testing a patch-embedding variant of VCMamba would separate that effect.
Share X Bluesky LinkedIn Reddit HN

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

3 major / 5 minor

Summary. The paper proposes VCMamba, a hierarchical vision backbone that combines convolutional feature extraction in early stages with multi-directional Mamba state-space blocks in the final stage. The architecture uses a convolutional stem, four stages with downsampling, and a final stage interleaving convolutional FFN blocks and Mamba blocks. Experiments on ImageNet-1K report VCMamba-B at 82.6% top-1 with 31.5M parameters, and on ADE20K semantic segmentation report 47.1 mIoU with Semantic FPN. The paper claims these results surpass PlainMamba-L3 and EfficientFormer-L7 with substantially fewer parameters. An ablation study shows incremental gains from skip connections, interleaving, LayerNorm, convolutional projections, and stage-wise batch normalization, totaling 2.4% over the authors' CNN-Mamba baseline.

Significance. If the reported comparisons hold, VCMamba offers a simple and effective hybrid of convolutional and state-space architectures, with the benefit of linear-complexity global modeling. The internal ablation is the strongest evidence: it quantifies the contribution of each architectural choice under a fixed training recipe. The release of code is a further strength. However, the headline claims of superiority over existing models rest on cross-paper comparisons with external published numbers, not on re-runs under VCMamba's own training protocol, and the reported margins are small (0.3% top-1, 2.0 mIoU). This makes the central comparative claims currently under-supported.

major comments (3)
  1. [§5.1, Table 2] The headline claims that VCMamba-B surpasses PlainMamba-L3 and ViG-B by 0.3% rely on accuracy numbers taken from the original papers (82.3% for both baselines). These models were trained under their own schedules, augmentations, and hyperparameters, none of which are re-run under VCMamba's 300-epoch AdamW/cosine/RandAugment/Mixup/CutMix recipe. A 0.3% margin is within typical seed-to-seed variation, and the paper reports no error bars or multiple seeds. This is load-bearing for the 'surpassing' claim. Please provide at least a re-benchmark of the key baselines under the same protocol, or report mean±std over multiple seeds for VCMamba and the baselines.
  2. [§5.2, Table 3] The ADE20K comparison against EfficientFormer-L7 (45.1 mIoU) uses the published number from [23] without confirming that the decoder, training iterations, input resolution, and pretraining are identical. The paper states VCMamba uses Semantic FPN with 40K iterations at 512×512, which may match EfficientFormer's protocol, but this is not explicitly verified. Since the reported margin is 2.0 mIoU and the parameter counts are backbone-only, a confounded protocol could erase the advantage. Please re-run the baseline under the same segmentation setup or provide the exact protocol of the cited number and justify comparability.
  3. [§5.3, Table 4] The ablation supports the incremental design decisions, but it does not test the central hybrid claim directly. The baseline already uses convolutional early stages and a PlainMamba-based final stage; the ablation never removes the convolutional early stages or replaces the Mamba stage with a fully convolutional or fully Mamba alternative. Thus the ablation validates block-level choices (e.g., LayerNorm gains +0.7%) rather than the value of the CNN+SSM combination itself. A comparison against a pure convolutional backbone (e.g., the same architecture with FFNs in all stages) or against the original PlainMamba-L3 under the same recipe would more directly support the stated contribution.
minor comments (5)
  1. [§4.2, Eq. (4)] Equation (4) is essentially identical to Equation (3) from PlainMamba, just with subscripts. It would be helpful to state explicitly what, if anything, is novel in VCMamba's direction-aware updating beyond adopting PlainMamba's mechanism.
  2. [§4.1, Figure 1] The FFN block description does not mention a residual connection, though Figure 1(d) appears to include one. Please clarify whether the FFN blocks use residual/additive paths, as this affects the architectural description.
  3. [Table 2] GMACs are missing for ViM-Ti and ViM-S. Consider adding them for completeness or note if they are unavailable.
  4. [§5.3, Table 4] The ablation reports a single run for each configuration. Reporting run-to-run variance (even for two seeds) would strengthen the conclusion that the 0.7% LayerNorm gain is significant rather than noise.
  5. [General] The paper provides a GitHub link but no commit hash, pretrained checkpoints, or training logs. Pinning a specific commit and releasing checkpoints would improve reproducibility, especially given the small margins claimed.

Circularity Check

0 steps flagged

No significant circularity: VCMamba's central claims are empirical comparisons against external published baselines, not derivations that reduce to their own inputs.

full rationale

VCMamba is an empirical architecture paper; it contains no 'prediction' or 'first-principles result' that could be equivalent to its inputs by construction. The headline margins (82.6% vs PlainMamba-L3 82.3%, ViG-B 82.3%; 47.1 vs EfficientFormer-L7 45.1) are cross-paper comparisons to externally published numbers, not to any quantity fitted from VCMamba's own pipeline. The paper does not define its architecture in terms of the target accuracy, nor does it fit a parameter to a subset and then call a closely related quantity a prediction. The Mamba-side components are explicitly attributed to an outside source: Section 4.2 states the block 'adapts the selective scan mechanism of Mamba [9] for 2D visual data by leveraging the scanning principles established in PlainMamba [47]' and the direction-aware update is 'as introduced in [47]' (Eq. 4); this is a borrowed component, not a self-cited uniqueness theorem. The ablation study (Table 4) starts from a 'Baseline (PlainMamba Stage 4)' and incrementally adds skip connection, interleaved FFN, LayerNorm, and conv heads; this is a standard empirical architecture search, and the final 82.6% is the measured result of that search, not a fitted constant renamed as a prediction. The paper's self-citations ([31,32,33]) appear in related work and standard training/segmentation recipes; they are not load-bearing for the central claim. The concern that the external baselines may have been trained under different recipes is a correctness/reproducibility risk, not circularity, and the paper reports no re-run of its baselines; however, by the circularity criteria (specific reduction to inputs or self-citation chain), no circular step can be exhibited.

Axiom & Free-Parameter Ledger

0 free parameters · 4 axioms · 0 invented entities

The central claims rest on standard deep-learning assumptions: that the baseline numbers from cited papers are accurate under their own protocols, that the PlainMamba 2D scan is a sound SSM adaptation, and that the inverted-residual FFN block is a strong local feature extractor. No new entities are introduced, so the invented_entities ledger is empty. No fitted constants are introduced.

axioms (4)
  • domain assumption The PlainMamba four-way Continuous 2D Scanning and direction-aware update provide a working SSM for 2D images.
    Section 4.2 adopts PlainMamba's scanning mechanism as the Mamba core without independent verification; any failure of that scan to capture 2D context would weaken the global-modeling stage.
  • domain assumption Inverted-residual-style FFN blocks (1x1 expand, 3x3 depthwise, 1x1 project) are an effective local feature extractor at early stages.
    Section 4.1 takes the MobileNetV2/EfficientFormer recipe as given; the paper does not ablate an alternative local block such as a standard residual conv.
  • domain assumption The reported values of existing models (PlainMamba-L3 82.3%, ViG-B 82.3%, EfficientFormer-L7 45.1 mIoU) are accurate and comparable.
    Used in Table 2 and Table 3 for the headline efficiency comparisons; the baseline numbers are cited rather than reproduced in the paper's pipeline.
  • domain assumption Cross-entropy training with AdamW, cosine schedule, and standard augmentations generalizes the same way across hybrid architectures.
    Section 5.1 relies on a standard training recipe to make cross-architecture accuracy comparisons meaningful.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of VCMamba: Bridging Convolutions with Multi-Directional Mamba for Efficient Visual Representation." pith.science (2026). https://pith.science/paper/FTVN7W7C

@misc{pith2026250904669,
  author       = {Pith},
  title        = {Pith review of: VCMamba: Bridging Convolutions with Multi-Directional Mamba for Efficient Visual Representation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FTVN7W7C}},
  note         = {Machine review of arXiv:2509.04669}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Recent advances in Vision Transformers (ViTs) and State Space Models (SSMs) have challenged the dominance of Convolutional Neural Networks (CNNs) in computer vision. ViTs excel at capturing global context, and SSMs like Mamba offer linear complexity for long sequences, yet they do not capture fine-grained local features as effectively as CNNs. Conversely, CNNs possess strong inductive biases for local features but lack the global reasoning capabilities of transformers and Mamba. To bridge this gap, we introduce \textit{VCMamba}, a novel vision backbone that integrates the strengths of CNNs and multi-directional Mamba SSMs. VCMamba employs a convolutional stem and a hierarchical structure with convolutional blocks in its early stages to extract rich local features. These convolutional blocks are then processed by later stages incorporating multi-directional Mamba blocks designed to efficiently model long-range dependencies and global context. This hybrid design allows for superior feature representation while maintaining linear complexity with respect to image resolution. We demonstrate VCMamba's effectiveness through extensive experiments on ImageNet-1K classification and ADE20K semantic segmentation. Our VCMamba-B achieves 82.6% top-1 accuracy on ImageNet-1K, surpassing PlainMamba-L3 by 0.3% with 37% fewer parameters, and outperforming Vision GNN-B by 0.3% with 64% fewer parameters. Furthermore, VCMamba-B obtains 47.1 mIoU on ADE20K, exceeding EfficientFormer-L7 by 2.0 mIoU while utilizing 62% fewer parameters. Code is available at https://github.com/Wertyuui345/VCMamba.

Figures

Figures reproduced from arXiv: 2509.04669 by Alex Zhang, Mustafa Munir, Radu Marculescu.

Figure 1
Figure 1. Figure 1: Overall architecture of VCMamba. An input image is passed through a convolutional head and four downsampling stages. The [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: The four multi-directional scanning patterns (e.g., Con [PITH_FULL_IMAGE:figures/full_fig_p005_2.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

54 extracted references · 31 canonical work pages · 2 internal anchors

  1. [1]

    Layer normalization.arXiv preprint arXiv:1607.06450,

    Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hin- ton. Layer normalization.arXiv preprint arXiv:1607.06450,

    Pith/arXiv arXiv

  2. [2]

    Mobile- former: Bridging mobilenet and transformer

    Yinpeng Chen, Xiyang Dai, Dongdong Chen, Mengchen Liu, Xiaoyi Dong, Lu Yuan, and Zicheng Liu. Mobile- former: Bridging mobilenet and transformer. InProceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5270–5279, 2022. 3

    work page 2022

  3. [3]

    PTQ4VM: Post-Training Quantization for Visual Mamba

    Younghyun Cho, Changhun Lee, Seonggon Kim, and Eun- hyeok Park. Ptq4vm: Post-training quantization for visual mamba.arXiv preprint arXiv:2412.20386, 2024. 3

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  4. [4]

    Randaugment: Practical automated data augmen- tation with a reduced search space

    Ekin D Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V Le. Randaugment: Practical automated data augmen- tation with a reduced search space. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 702–703, 2020. 6

    work page 2020

  5. [5]

    Coatnet: Marrying convolution and attention for all data sizes.Advances in neural information processing systems, 34:3965–3977, 2021

    Zihang Dai, Hanxiao Liu, Quoc V Le, and Mingxing Tan. Coatnet: Marrying convolution and attention for all data sizes.Advances in neural information processing systems, 34:3965–3977, 2021. 3

    work page 2021

  6. [6]

    Imagenet: A large-scale hierarchical image database

    Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248–255, 2009. 2, 6

    work page 2009

  7. [7]

    An image is worth 16x16 words: Transformers for image recognition at scale.arXiv preprint arXiv:2010.11929, 2020

    Alexey Dosovitskiy et al. An image is worth 16x16 words: Transformers for image recognition at scale.arXiv preprint arXiv:2010.11929, 2020. 1, 2

    Pith/arXiv arXiv 2010

  8. [8]

    Sigmoid- weighted linear units for neural network function approxima- tion in reinforcement learning.Neural networks, 107:3–11,

    Stefan Elfwing, Eiji Uchibe, and Kenji Doya. Sigmoid- weighted linear units for neural network function approxima- tion in reinforcement learning.Neural networks, 107:3–11,

  9. [9]

    Mamba: Linear-time sequence modeling with selective state spaces.arXiv preprint arXiv:2312.00752, 2023

    Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces.arXiv preprint arXiv:2312.00752, 2023. 1, 2, 3, 5

    Pith/arXiv arXiv 2023

  10. [10]

    Q-mamba: On first exploration of vision mamba for image quality assessment.arXiv preprint arXiv:2406.09546, 2024

    Fengbin Guan, Xin Li, Zihao Yu, Yiting Lu, and Zhibo Chen. Q-mamba: On first exploration of vision mamba for image quality assessment.arXiv preprint arXiv:2406.09546, 2024. 3

    Pith/arXiv arXiv 2024

  11. [11]

    Demystify mamba in vision: A linear attention perspective.arXiv preprint arXiv:2405.16605, 2024

    Dongchen Han, Ziyi Wang, Zhuofan Xia, Yizeng Han, Yi- fan Pu, Chunjiang Ge, Jun Song, Shiji Song, Bo Zheng, and Gao Huang. Demystify mamba in vision: A linear attention perspective.arXiv preprint arXiv:2405.16605, 2024. 3

    Pith/arXiv arXiv 2024

  12. [12]

    Vision gnn: An image is worth graph of nodes

    Kai Han, Yunhe Wang, Jianyuan Guo, Yehui Tang, and En- hua Wu. Vision gnn: An image is worth graph of nodes. arXiv preprint arXiv:2206.00272, 2022. 2, 3, 6, 7

    Pith/arXiv arXiv 2022

  13. [13]

    Mambavision: A hy- brid mamba-transformer vision backbone.arXiv preprint arXiv:2407.08083, 2024

    Ali Hatamizadeh and Jan Kautz. Mambavision: A hy- brid mamba-transformer vision backbone.arXiv preprint arXiv:2407.08083, 2024. 3, 7

    Pith/arXiv arXiv 2024

  14. [14]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. InProceed- ings of the IEEE conference on computer vision and pattern recognition, pages 770–778, 2016. 1, 2, 4, 7

    work page 2016

  15. [15]

    Gaussian error linear units (gelus).arXiv preprint arXiv:1606.08415, 2016

    Dan Hendrycks and Kevin Gimpel. Gaussian error linear units (gelus).arXiv preprint arXiv:1606.08415, 2016. 4, 5

    Pith/arXiv arXiv 2016

  16. [16]

    Mobilenets: Efficient convolu- tional neural networks for mobile vision applications.arXiv preprint arXiv:1704.04861, 2017

    Andrew G Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco An- dreetto, and Hartwig Adam. Mobilenets: Efficient convolu- tional neural networks for mobile vision applications.arXiv preprint arXiv:1704.04861, 2017. 2, 4

    Pith/arXiv arXiv 2017

  17. [17]

    Batch normalization: Accelerating deep network training by reducing internal co- variate shift

    Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal co- variate shift. InInternational conference on machine learn- ing, pages 448–456. pmlr, 2015. 3, 4

    work page 2015

  18. [18]

    Panoptic feature pyramid networks

    Alexander Kirillov, Ross Girshick, Kaiming He, and Piotr Doll´ar. Panoptic feature pyramid networks. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6399–6408, 2019. 6

    work page 2019

  19. [19]

    Imagenet classification with deep convolutional neural net- works.Advances in neural information processing systems, 25, 2012

    Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural net- works.Advances in neural information processing systems, 25, 2012. 1, 2

    work page 2012

  20. [20]

    Gradient-based learning applied to document recog- nition.Proceedings of the IEEE, 86(11):2278–2324, 1998

    Yann LeCun, L ´eon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to document recog- nition.Proceedings of the IEEE, 86(11):2278–2324, 1998. 1, 2

    work page 1998

  21. [21]

    Videomamba: State space model for efficient video understanding

    Kunchang Li, Xinhao Li, Yi Wang, Yinan He, Yali Wang, Limin Wang, and Yu Qiao. Videomamba: State space model for efficient video understanding. InEuropean Conference on Computer Vision, pages 237–255. Springer, 2024. 3

    work page 2024

  22. [22]

    Rethinking Vision Transformers for MobileNet Size and Speed

    Yanyu Li, Ju Hu, Yang Wen, Georgios Evangelidis, Kamyar Salahi, Yanzhi Wang, Sergey Tulyakov, and Jian Ren. Re- thinking vision transformers for mobilenet size and speed. arXiv preprint arXiv:2212.08059, 2022. 3, 6

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  23. [23]

    Effi- cientformer: Vision transformers at mobilenet speed.arXiv preprint arXiv:2206.01191, 2022

    Yanyu Li, Geng Yuan, Yang Wen, Eric Hu, Georgios Evan- gelidis, Sergey Tulyakov, Yanzhi Wang, and Jian Ren. Effi- cientformer: Vision transformers at mobilenet speed.arXiv preprint arXiv:2206.01191, 2022. 2, 3, 6, 7

    Pith/arXiv arXiv 2022

  24. [24]

    Vision mamba: A comprehensive survey and taxonomy.arXiv preprint arXiv:2405.04404, 2024

    Xiao Liu, Chenxu Zhang, and Lei Zhang. Vision mamba: A comprehensive survey and taxonomy.arXiv preprint arXiv:2405.04404, 2024. 2 8

    Pith/arXiv arXiv 2024

  25. [25]

    Vmamba: Visual state space model.Advances in neural information processing systems, 37:103031–103063, 2024

    Yue Liu, Yunjie Tian, Yuzhong Zhao, Hongtian Yu, Lingxi Xie, Yaowei Wang, Qixiang Ye, Jianbin Jiao, and Yunfan Liu. Vmamba: Visual state space model.Advances in neural information processing systems, 37:103031–103063, 2024. 1, 2

    work page 2024

  26. [26]

    Swin transformer: Hierarchical vision transformer using shifted windows

    Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012–10022, 2021. 2, 4

    2021

  27. [27]

    A convnet for the 2020s

    Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feicht- enhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. InProceedings of the IEEE/CVF conference on com- puter vision and pattern recognition, pages 11976–11986,

  28. [28]

    Decoupled weight decay regularization.arXiv preprint arXiv:1711.05101, 2017

    Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization.arXiv preprint arXiv:1711.05101, 2017. 6

    Pith/arXiv arXiv 2017

  29. [29]

    Mobilevit: light- weight, general-purpose, and mobile-friendly vision trans- former.arXiv preprint arXiv:2110.02178, 2021

    Sachin Mehta and Mohammad Rastegari. Mobilevit: light- weight, general-purpose, and mobile-friendly vision trans- former.arXiv preprint arXiv:2110.02178, 2021. 2

    Pith/arXiv arXiv 2021

  30. [30]

    Separable self- attention for mobile vision transformers.arXiv preprint arXiv:2206.02680, 2022

    Sachin Mehta and Mohammad Rastegari. Separable self- attention for mobile vision transformers.arXiv preprint arXiv:2206.02680, 2022. 2

    Pith/arXiv arXiv 2022

  31. [31]

    Mo- bilevig: Graph-based sparse attention for mobile vision ap- plications

    Mustafa Munir, William Avery, and Radu Marculescu. Mo- bilevig: Graph-based sparse attention for mobile vision ap- plications. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Work- shops, pages 2211–2219, 2023. 2, 6

    work page 2023

  32. [32]

    Greedyvig: Dynamic axial graph construction for efficient vision gnns

    Mustafa Munir, William Avery, Md Mostafijur Rahman, and Radu Marculescu. Greedyvig: Dynamic axial graph construction for efficient vision gnns. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6118–6127, 2024. 2, 6

    work page 2024

  33. [33]

    Rapidnet: Multi-level dilated convolution based mobile backbone

    Mustafa Munir, Md Mostafijur Rahman, and Radu Mar- culescu. Rapidnet: Multi-level dilated convolution based mobile backbone. In2025 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 8302–8312. IEEE, 2025. 2, 4, 6, 7

    work page 2025

  34. [34]

    Rectified linear units im- prove restricted boltzmann machines

    Vinod Nair and Geoffrey E Hinton. Rectified linear units im- prove restricted boltzmann machines. InProceedings of the 27th international conference on machine learning (ICML- 10), pages 807–814, 2010. 3

    work page 2010

  35. [35]

    Clustervig: Efficient globally aware vision gnns via image partitioning.arXiv preprint arXiv:2501.10640, 2025

    Dhruv Parikh, Jacob Fein-Ashley, Tian Ye, Rajgopal Kan- nan, and Viktor Prasanna. Clustervig: Efficient globally aware vision gnns via image partitioning.arXiv preprint arXiv:2501.10640, 2025. 2

    Pith/arXiv arXiv 2025

  36. [36]

    Pytorch: An imperative style, high- performance deep learning library.Advances in neural in- formation processing systems, 32, 2019

    Adam Paszke et al. Pytorch: An imperative style, high- performance deep learning library.Advances in neural in- formation processing systems, 32, 2019. 6

    work page 2019

  37. [37]

    Mobilenetv2: Inverted residuals and linear bottlenecks

    Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zh- moginov, and Liang-Chieh Chen. Mobilenetv2: Inverted residuals and linear bottlenecks. InProceedings of the IEEE conference on computer vision and pattern recogni- tion, pages 4510–4520, 2018. 2, 4

    work page 2018

  38. [38]

    Very deep convo- lutional networks for large-scale image recognition.arXiv preprint arXiv:1409.1556, 2014

    Karen Simonyan and Andrew Zisserman. Very deep convo- lutional networks for large-scale image recognition.arXiv preprint arXiv:1409.1556, 2014. 2

    Pith/arXiv arXiv 2014

  39. [39]

    Wignet: Windowed vi- sion graph neural network

    Gabriele Spadaro, Marco Grangetto, Attilio Fiandrotti, Enzo Tartaglione, and Jhony H Giraldo. Wignet: Windowed vi- sion graph neural network. In2025 IEEE/CVF Winter Con- ference on Applications of Computer Vision (WACV), pages 859–868. IEEE, 2025. 2

    work page 2025

  40. [40]

    Efficientnet: Rethinking model scaling for convolutional neural networks

    Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. InInternational conference on machine learning, pages 6105–6114. PMLR,

  41. [41]

    Training data-efficient image transformers & distillation through at- tention

    Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Herv ´e J´egou. Training data-efficient image transformers & distillation through at- tention. InInternational conference on machine learning, pages 10347–10357. PMLR, 2021. 2, 6, 7

    work page 2021

  42. [42]

    Fastvit: A fast hybrid vision transformer using structural reparameterization

    Pavan Kumar Anasosalu Vasu, James Gabriel, Jeff Zhu, On- cel Tuzel, and Anurag Ranjan. Fastvit: A fast hybrid vision transformer using structural reparameterization. InProceed- ings of the IEEE/CVF International Conference on Com- puter Vision, 2023. 7

    work page 2023

  43. [43]

    Attention is all you need.Advances in neural information processing systems, 30, 2017

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszko- reit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need.Advances in neural information processing systems, 30, 2017. 1, 2

    work page 2017

  44. [44]

    Repvit: Revisiting mobile cnn from vit perspective

    Ao Wang, Hui Chen, Zijia Lin, Jungong Han, and Guiguang Ding. Repvit: Revisiting mobile cnn from vit perspective. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15909–15920, 2024. 2

    work page 2024

  45. [45]

    Pyramid vision transformer: A versatile backbone for dense prediction without convolutions

    Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling 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. 2, 3, 6, 7

    work page 2021

  46. [46]

    PyTorch Image Models.https : / / github

    Ross Wightman. PyTorch Image Models.https : / / github . com / rwightman / pytorch - image - models, 2019. 6

    work page 2019

  47. [47]

    Plainmamba: Improving non-hierarchical mamba in visual recognition.arXiv preprint arXiv:2403.17695, 2024

    Chenhongyi Yang, Zehui Chen, Miguel Espinosa, Linus Er- icsson, Zhenyu Wang, Jiaming Liu, and Elliot J Crowley. Plainmamba: Improving non-hierarchical mamba in visual recognition.arXiv preprint arXiv:2403.17695, 2024. 1, 2, 3, 5, 6, 7

    Pith/arXiv arXiv 2024

  48. [48]

    Mambaout: Do we really need mamba for vision?arXiv preprint arXiv:2405.07992,

    Weihao Yu and Xinchao Wang. Mambaout: Do we really need mamba for vision?arXiv preprint arXiv:2405.07992,

    Pith/arXiv arXiv

  49. [49]

    Metaformer is actually what you need for vision

    Weihao Yu, Mi Luo, Pan Zhou, Chenyang Si, Yichen Zhou, Xinchao Wang, Jiashi Feng, and Shuicheng Yan. Metaformer is actually what you need for vision. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10819–10829, 2022. 6, 7

    work page 2022

  50. [50]

    Cutmix: Regu- larization strategy to train strong classifiers with localizable features

    Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regu- larization strategy to train strong classifiers with localizable features. InProceedings of the IEEE/CVF international con- ference on computer vision, pages 6023–6032, 2019. 6

    work page 2019

  51. [51]

    Dauphin, and David Lopez-Paz

    Hongyi Zhang, Moustapha Cisse, Yann N. Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimiza- 9 tion. InInternational Conference on Learning Representa- tions, 2018. 6

    work page 2018

  52. [52]

    Random erasing data augmentation

    Zhun Zhong, Liang Zheng, Guoliang Kang, Shaozi Li, and Yi Yang. Random erasing data augmentation. InProceed- ings of the AAAI conference on artificial intelligence, pages 13001–13008, 2020. 6

    work page 2020

  53. [53]

    Scene parsing through ade20k dataset

    Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 633–641,

  54. [54]

    Vision mamba: Efficient visual representation learning with bidirectional state space model

    Lianghui Zhu, Bencheng Liao, Qian Zhang, Xinlong Wang, Wenyu Liu, and Xinggang Wang. Vision mamba: Efficient visual representation learning with bidirectional state space model. InForty-first International Conference on Machine Learning. 1, 2, 3, 6, 7 10

    work page

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.