Pith. sign in

REVIEW 3 major objections 4 minor 79 references

A2Mamba: Attention-augmented State Space Models for Visual Recognition

T0 review · 3 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read A hybrid token mixer that feeds multi-scale attention maps into the hidden states of a state space model achieves 86.1% top-1 accuracy on ImageNet-1K.

desk verdict A genuinely new attention-SSM interaction backed by solid ablations, but the universal SOTA claim outruns the evidence. read the letter →

arxiv 2507.16624 v1 pith:O3SK6OJP submitted 2025-07-22 cs.CV

classification cs.CV
keywords visualrecognitionstatespacemodelsMambahybridvisionbackbonemulti-scaleattentionattention-augmentedSSMImageNetclassificationsemanticsegmentation
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper aims to establish that attention and state space models can be fused within a single token mixer, rather than simply stacked, and that this fusion yields a better vision backbone than either family alone. It introduces A2Mamba, a four-stage pyramid network whose core mixer MASS computes multi-scale attention maps and then feeds those maps into an attention-augmented state space model (A2SSM). By spatially aggregating the SSM's hidden states with attention maps, the mixer gives the sequential scan a two-dimensional spatial sense and reduces the causality that pure scanning imposes. The paper reports that A2Mamba-L reaches 86.1% top-1 accuracy on ImageNet-1K and that smaller variants beat prior ConvNet-, Transformer-, and Mamba-based backbones on classification, segmentation, and detection, often at lower compute. If these results hold, the architecture answers a concrete design question: deep cross-attention between attention maps and SSM hidden states is a more effective hybrid recipe than layer stacking.

What carries the argument

The load-bearing object is MASS, the Multi-scale Attention-augmented State Space Model, a token mixer. Its first branch is sliding local attention (SLA), which computes an attention map for each token over a fixed window; its second branch is dilated sliding attention (DLA), which uses a dilation rate chosen so the dilated window spans the whole feature map, producing a second attention map. Inside A2SSM, the selective scan's hidden states are split and aggregated with those two attention maps, acting as learned spatial pooling over local and long-range neighborhoods before the states modulate the SSM output. That shared use of attention maps is the mechanism that converts a one-dimensional causal scan into a two-dimensionally aware representation without quadratic complexity.

What would settle it

Take A2Mamba-S and its closest comparators, for example RMT-S, TransNeXt-T, and MambaVision-B, and retrain all of them from the same ImageNet-1K schedule, augmentation, and fine-tuning recipe; if the top-1 gap of roughly 0.6 to 1.0 points shrinks or reverses under identical conditions, the universal claim of outperforming all prior backbones fails. The same check applies to the ADE20K mIoU margins with UperNet and the COCO AP margins with Cascade Mask R-CNN.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that hybrid Transformer-Mamba backbones have been under-built: stacking attention layers beside SSM layers leaves their representations only loosely coupled. A2Mamba replaces that with a single MASS token mixer, in which sliding local attention and dilated sliding attention generate multi-scale attention maps for the input, and those same maps are applied as a cross-attention-style aggregation to the hidden states produced by a selective scan SSM. The resulting hidden states are then multiplied with the SSM's output projection, so global context modulation is carried out on attention-enhanced states. The paper reports this design outperforms all previous ConvNet-, Transformer-, and Mamba-based architectures in every task evaluated, with specific margins such as 86.1% top-1 on ImageNet-1K, a 2.5% mIoU gain over CAFormer-S36 in ADE20K segmentation, and a 1.2%/0.9% APb/APm gain over MambaVision-B with 40% fewer parameters in Cascade Mask R-CNN.

Load-bearing premise

The load-bearing premise is that the published numbers for all the baselines in Tables 2, 4, 5, and 6 are accurate, up to date, and produced under training and evaluation protocols comparable to A2Mamba, since the paper does not re-run every baseline under identical conditions.

Editorial extensions

If this is right

  • If the reported numbers are correct, A2Mamba sets a new state of the art for hybrid vision backbones in ImageNet classification, COCO detection and instance segmentation, and ADE20K semantic segmentation at several model scales.
  • Because the mixer keeps linear-time modules, A2Mamba can be applied to high-resolution inputs where Transformer baselines incur growing memory and latency; the paper's 1024x1024 measurements show less accuracy loss per extra compute.
  • A single-scan SSM augmented by attention maps matches or exceeds multi-direction scanning baselines, suggesting future Mamba vision designs can simplify scan-direction engineering.
  • The MM-Refine decoder and SegMAN-V2 show the same MASS mixer works for dense prediction, so the integration idea extends beyond the backbone itself.

Reading between the lines

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

  • Editorial inference: if multi-scale attention maps act as trainable spatial priors for the hidden state, the same cross-attention scheme could transfer to other structured sequence domains, such as video or point clouds, where sequential scans also fight two- or three-dimensional structure.
  • Editorial inference: the adaptive dilation rule ties the effective attention coverage to input resolution, so the model is plausibly more robust to resolution shifts than fixed-window attention; that property could be tested on out-of-distribution resolutions without fine-tuning.
  • Editorial inference: the reported efficiency gains at high resolution suggest the architecture may suit deployment on edge hardware, but an independent benchmark under identical schedules and hardware is needed before treating those margins as universal.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. This manuscript extends the authors' CVPR 2025 SegMAN paper by proposing A2Mamba, a hierarchical hybrid vision backbone whose core token mixer, MASS, combines multi-scale sliding/dilated local attention with a selective state space model. In A2SSM, the precomputed attention maps are used to spatially aggregate the SSM hidden states before the output projection, which the authors argue improves 2D spatial coherence and reduces the causality introduced by sequential scanning. The paper reports five model sizes, an upgraded segmentation decoder (MM-Refine / SegMAN-V2), and experiments on ImageNet-1K classification, COCO object detection and instance segmentation, ADE20K/Cityscapes/COCO-Stuff semantic segmentation, plus speed and memory measurements at increasing resolutions. The headline claim is that A2Mamba outperforms all previous ConvNet-, Transformer-, and Mamba-based architectures; for example, A2Mamba-L reaches 86.1% ImageNet-1K top-1 accuracy.

Significance. Assuming the reported numbers are accurate and protocol-comparable, A2Mamba is a credible architectural contribution: it introduces a non-trivial interaction between multi-scale attention maps and SSM hidden states, achieves competitive accuracy with linear-time long-range modeling, and is accompanied by a public code release and detailed ablation roadmaps (Tables 9-13). The ablations are generally informative and support the central design choices. The main limitation is that the universal state-of-the-art claim is not verified under controlled conditions: the reported margins over several closest baselines are small (around 0.4-0.7% in Table 2), and the comparisons rely on published numbers from heterogeneous training and evaluation protocols rather than re-runs. This makes the strongest claim load-bearing on an assumption that is not tested in the paper.

major comments (3)
  1. [§4, Tables 2, 4-7] The claim that A2Mamba "outperforms all previous ConvNet-, Transformer-, and Mamba-based architectures" rests entirely on baseline numbers taken from other papers, with different training schedules, augmentation choices, stochastic depth rates, decoder designs, and evaluation conventions. The reported margins are small; for example, Table 2 gives A2Mamba-S vs. RMT-S as +0.6%, A2Mamba-B vs. MLLA-B as +0.4%, and A2Mamba-L vs. CAFormer-B36 as +0.6%. Since typical protocol differences can shift top-1 accuracy by more than these margins, the unqualified universal claim is not established by the submitted evidence. I ask the authors to either re-run the closest baselines under an identical protocol or to explicitly rephrase the claim so that it refers to the reported numbers rather than an unconditional superiority statement.
  2. [§4.4, Table 8] The 512x512 and 1024x1024 columns evaluate models pre-trained at 224x224 without fine-tuning, i.e., zero-shot resolution transfer. This is not a standard evaluation protocol, and the large accuracy differences at high resolution may reflect positional-encoding behavior, interpolation artifacts, or scanning-order assumptions rather than intrinsic superiority of the architecture. Calling these results "stronger generalization ability" is misleading unless the protocol is clearly labeled as zero-shot transfer and the comparison is made against baselines under exactly the same conditions.
  3. [Tables 2 and 8] There is a direct inconsistency in the headline numbers: Table 2 lists A2Mamba-L as 95M parameters with 86.1% accuracy, while Table 8 lists the same model as 94M parameters with 86.2% accuracy. Because the paper's central claim is empirical and quantitative, this discrepancy must be resolved or explicitly explained; otherwise it undermines confidence in the other tabulated values.
minor comments (4)
  1. [§3.3, Eq. (2)] Equation (2) writes "A2, X'_2 = DLA(X1)", but the surrounding text states that X2 is processed by DLA and X1 by SLA. This typo should be corrected to DLA(X2), since the current formula is inconsistent with the textual description and with Figure 3.
  2. [Table 3] TransNeXt-S and TransNeXt-B in Table 3 are cited as [46], but reference [46] is TransXNet; the correct citation appears to be [26] (TransNeXt). This should be fixed.
  3. [§3.3] The sentence "Note that d_state is omitted because previous work [15], [19], [20], [27] set it to 1" is unclear to readers who do not know the Mamba implementation; please define d_state or give the exact selective-scan formulation used in A2SSM.
  4. [Tables 9 and 10] The baseline row in Table 10 reports 83.0% accuracy and 49.7% mIoU for the same configuration that Table 9 lists as "w Gate" with 82.9% accuracy and 49.7% mIoU. These duplicated numbers should be reconciled.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: A2Mamba's central performance claims are measured against external benchmarks and are not derived from its inputs or from self-citations.

full rationale

The paper's central claim is that A2Mamba outperforms prior ConvNet-, Transformer-, and Mamba-based models on ImageNet-1K, COCO detection and instance segmentation, and ADE20K semantic segmentation. These claims are evaluated on public datasets with independently reported baselines; the architecture's MASS mixer is specified as a combination of sliding and dilated attention modules and an SSM, with no parameter fitted to the reported accuracies and no equation that constructs the target result from the inputs. Self-citations to the authors' prior SegMAN, SparX, TransXNet, and OverLoCK appear as component lineage and as comparison baselines, but none of these is invoked as an external authority that forces the outcome; there is no uniqueness theorem, no ansatz smuggled in via citation, and no fitted-input-called-prediction pattern. The ablation 'roadmap' compares variants on external validation metrics and even reports negative results (e.g., fixed dilation rates and alternate SSM mixers), which shows the claimed improvements are empirically falsifiable rather than true by construction. Concerns about protocol comparability of cited baselines (e.g., heterogeneous training schedules, small 0.6% margins in Table 2, and Table 8's off-resolution inference) are correctness risks, not circularity, because the paper does not define its superiority into existence. Therefore no circular step meeting the required evidence standard is present.

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

The central claim is empirical and does not rest on a mathematical derivation. It depends on the correctness of standard SSM equations (taken from Mamba), on the architectural assumption that attention maps transfer to hidden-state aggregation, and on the comparability of baseline results. No physical entities are introduced; the paper proposes algorithmic modules. Free parameters are the hand-set window sizes, model widths/depths, and the dilation design choice.

free parameters (3)
  • Sliding window sizes (K) per stage = [11, 9, 7, 7] for stages 1-4
    Set following prior work SegMAN [27]; controls the receptive field of the local attention branches. These are hand-chosen architectural hyperparameters, not learned.
  • Model architecture configuration (channels, depths, heads per stage) = See Table 1 for variants N/T/S/B/L
    Standard practice; these define model capacity and are chosen to match comparable baselines. They are hand-set, not fitted to the result.
  • Dilation rate formula parameter K = K is the window size (above), and r=(int(H/K), int(W/K))
    The adaptive dilation rate depends on input resolution; it is a design choice rather than a fitted constant.
assumptions (3)
  • standard math State space model formulation from Mamba (Gu & Dao, 2024) is assumed correct and used as a building block.
    The paper relies on the selective scan equations and hidden state recurrence without re-deriving them (Section 3.3).
  • domain assumption Attention maps computed on input features remain valid and useful for aggregating SSM hidden states.
    A2SSM assumes that the attention affinities from X are transferable to the hidden state sequence S; this is empirically validated but not theoretically justified.
  • domain assumption The training and evaluation protocols for all baselines are assumed to be directly comparable.
    The claim of outperforming all previous architectures depends on the comparability of numbers taken from prior papers, which may use different schedules, resolutions, or decoder settings.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A2Mamba: Attention-augmented State Space Models for Visual Recognition." pith.science (2026). https://pith.science/paper/O3SK6OJP

@misc{pith2026250716624,
  author       = {Pith},
  title        = {Pith review of: A2Mamba: Attention-augmented State Space Models for Visual Recognition},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/O3SK6OJP}},
  note         = {Machine review of arXiv:2507.16624}
}
read the original abstract

Transformers and Mamba, initially invented for natural language processing, have inspired backbone architectures for visual recognition. Recent studies integrated Local Attention Transformers with Mamba to capture both local details and global contexts. Despite competitive performance, these methods are limited to simple stacking of Transformer and Mamba layers without any interaction mechanism between them. Thus, deep integration between Transformer and Mamba layers remains an open problem. We address this problem by proposing A2Mamba, a powerful Transformer-Mamba hybrid network architecture, featuring a new token mixer termed Multi-scale Attention-augmented State Space Model (MASS), where multi-scale attention maps are integrated into an attention-augmented SSM (A2SSM). A key step of A2SSM performs a variant of cross-attention by spatially aggregating the SSM's hidden states using the multi-scale attention maps, which enhances spatial dependencies pertaining to a two-dimensional space while improving the dynamic modeling capabilities of SSMs. Our A2Mamba outperforms all previous ConvNet-, Transformer-, and Mamba-based architectures in visual recognition tasks. For instance, A2Mamba-L achieves an impressive 86.1% top-1 accuracy on ImageNet-1K. In semantic segmentation, A2Mamba-B exceeds CAFormer-S36 by 2.5% in mIoU, while exhibiting higher efficiency. In object detection and instance segmentation with Cascade Mask R-CNN, A2Mamba-S surpasses MambaVision-B by 1.2%/0.9% in AP^b/AP^m, while having 40% less parameters. Code is publicly available at https://github.com/LMMMEng/A2Mamba.

Figures

Figures reproduced from arXiv: 2507.16624 by the authors.

Figure 1
Figure 1. Performance comparisons between our A2Mamba and other [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. The overall architecture of the proposed A2Mamba. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Workflow of our MASS token mixer. Pixel Unshuffle Pixel Unshuffle 3 × 𝟑 Conv S=2 MASS ConvFFN 𝟑𝑪 × 𝑯 𝟑𝟐 × 𝑾 𝟑𝟐 𝟒𝑪 × 𝑯 𝟖 × 𝑾 𝟖 3 × 𝟑 Conv S=2 5 × 5 RepConv UP Concat 𝑪 × 𝑯 𝟑𝟐 × 𝑾 𝟑𝟐 𝑪 × 𝑯 𝟑𝟐 × 𝑾 𝟑𝟐 Concat Concat GAP MM-Refine 1 × 1 Conv 1 × 1 Conv 1 × 1 Conv 1 × 1 Conv UP UP 1 ×1 Conv UP 1 ×1 Conv UP 𝑪𝟒 × 𝑯 𝟑𝟐 × 𝑾 𝟑𝟐 𝑪𝟑 × 𝑯 𝟏𝟔 × 𝑾 𝟏𝟔 𝑪𝟐 × 𝑯 𝟖 × 𝑾 𝟖 𝑪𝟏 × 𝑯 𝟒 × 𝑾 𝟒 Stage 1 Stage 2 Stage 3 Stage 4 𝐅 Concat UP 𝐅𝒍 1 ×1 Co… view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: The overall architecture of SegMAN-V2. and C ′ . Then, ∆, B, and a learnable vector A are used to generate a causal hidden state map (HSM) S ∈ R C×HW through the selective scan operation, where the t-th token dynamically fuses the tokens at all previous positions. Note…
Figure 5
Figure 5. Figure 5: Comparison of ERF among various models. in [PITH_FULL_IMAGE:figures/full_fig_p010_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

79 extracted references · 74 canonical work pages

  1. [1]

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

    A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gelly, et al. , “An image is worth 16x16 words: Transformers for image recognition at scale,” in International Conference on Learning Repre- sentations, 2021

  2. [2]

    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 Information Processing Systems , vol. 30, 2017

  3. [3]

    Swin transformer: Hierarchical vision transformer using shifted windows,

    Z. Liu, Y. Lin, Y. Cao, H. Hu, Y. Wei, Z. Zhang, S. Lin, and B. Guo, “Swin transformer: Hierarchical vision transformer using shifted windows,” in Proceedings of the IEEE/CVF International Conference on Computer Vision, pp. 10012–10022, 2021

  4. [4]

    Cswin transformer: A general vision transformer back- bone with cross-shaped windows,

    X. Dong, J. Bao, D. Chen, W. Zhang, N. Yu, L. Yuan, D. Chen, and B. Guo, “Cswin transformer: A general vision transformer back- bone with cross-shaped windows,” in Proceedings of the IEEE/CVF International Conference on Computer Vision and Pattern Recognition , pp. 12124–12134, 2022

  5. [5]

    Neighborhood attention transformer,

    A. Hassani, S. Walton, J. Li, S. Li, and H. Shi, “Neighborhood attention transformer,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 6185–6194, 2023

  6. [6]

    Biformer: Vi- sion transformer with bi-level routing attention,

    L. Zhu, X. Wang, Z. Ke, W. Zhang, and R. W. Lau, “Biformer: Vi- sion transformer with bi-level routing attention,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 10323–10333, 2023

  7. [7]

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

    W. Wang, E. Xie, X. Li, D.-P . Fan, K. Song, D. Liang, T. Lu, P . Luo, and L. Shao, “Pyramid vision transformer: A versatile backbone for dense prediction without convolutions,” in Proceedings of the IEEE/CVF Conference on Computer Vision, pp. 568–578, 2021

  8. [8]

    Pvt v2: Improved baselines with pyramid vision transformer,

    W. Wang, E. Xie, X. Li, D.-P . Fan, K. Song, D. Liang, T. Lu, P . Luo, and L. Shao, “Pvt v2: Improved baselines with pyramid vision transformer,” Computational Visual Media , vol. 8, no. 3, pp. 415– 424, 2022

Show all 79 references
  1. [9]

    P2t: Pyramid pool- ing transformer for scene understanding,

    Y.-H. Wu, Y. Liu, X. Zhan, and M.-M. Cheng, “P2t: Pyramid pool- ing transformer for scene understanding,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 45, no. 11, pp. 12760– 12771, 2023

  2. [10]

    Maxvit: Multi-axis vision transformer,

    Z. Tu, H. Talebi, H. Zhang, F. Yang, P . Milanfar, A. Bovik, and Y. Li, “Maxvit: Multi-axis vision transformer,” in European Conference on Computer Vision, pp. 459–479, Springer, 2022

  3. [11]

    Dilated neighborhood attention trans- former,

    A. Hassani and H. Shi, “Dilated neighborhood attention trans- former,” arXiv preprint arXiv:2209.15001, 2022

  4. [12]

    Crossformer++: A versatile vision transformer hinging on cross-scale attention,

    W. Wang, W. Chen, Q. Qiu, L. Chen, B. Wu, B. Lin, X. He, and W. Liu, “Crossformer++: A versatile vision transformer hinging on cross-scale attention,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 46, no. 5, pp. 3123–3136, 2023

  5. [13]

    Mamba: Linear-time sequence modeling with selective state spaces,

    A. Gu and T. Dao, “Mamba: Linear-time sequence modeling with selective state spaces,” in Conference on Language Modeling, 2024

  6. [14]

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

    L. Zhu, B. Liao, Q. Zhang, X. Wang, W. Liu, and X. Wang, “Vision mamba: Efficient visual representation learning with bidirectional state space model,” in International Conference on Machine Learning, vol. 235, pp. 62429–62442, 2024

  7. [15]

    Vmamba: Visual state space model,

    Y. Liu, Y. Tian, Y. Zhao, H. Yu, L. Xie, Y. Wang, Q. Ye, J. Jiao, and Y. Liu, “Vmamba: Visual state space model,” Advances in neural information processing systems, vol. 37, pp. 103031–103063, 2024

  8. [16]

    Local- mamba: Visual state space model with windowed selective scan,

    T. Huang, X. Pei, S. You, F. Wang, C. Qian, and C. Xu, “Local- mamba: Visual state space model with windowed selective scan,” in European Conference on Computer Vision Workshop, 2024

  9. [17]

    Plainmamba: Improving non-hierarchical mamba in visual recognition,

    C. Yang, Z. Chen, M. Espinosa, L. Ericsson, Z. Wang, J. Liu, and E. J. Crowley, “Plainmamba: Improving non-hierarchical mamba in visual recognition,” in British Machine Vision Conference, 2024

  10. [18]

    Efficientvmamba: Atrous selective scan for light weight visual mamba,

    X. Pei, T. Huang, and C. Xu, “Efficientvmamba: Atrous selective scan for light weight visual mamba,” in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 39, pp. 6443–6451, 2025

  11. [19]

    Spatial- mamba: Effective visual state space models via structure-aware state fusion,

    C. Xiao, M. Li, Z. Zhang, D. Meng, and L. Zhang, “Spatial- mamba: Effective visual state space models via structure-aware state fusion,” in International Conference on Learning Representations, 2025

  12. [20]

    Sparx: A sparse cross-layer connec- tion mechanism for hierarchical vision mamba and transformer networks,

    M. Lou, Y. Fu, and Y. Yu, “Sparx: A sparse cross-layer connec- tion mechanism for hierarchical vision mamba and transformer networks,” in Proceedings of the AAAI Conference on Artificial Intel- ligence, vol. 39, pp. 19104–19114, 2025

  13. [21]

    Mambaout: Do we really need mamba for vision?,

    W. Yu and X. Wang, “Mambaout: Do we really need mamba for vision?,” in Proceedings of the Computer Vision and Pattern Recognition Conference, pp. 4484–4496, 2025

  14. [22]

    Mambavision: A hybrid mamba- transformer vision backbone,

    A. Hatamizadeh and J. Kautz, “Mambavision: A hybrid mamba- transformer vision backbone,” in Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, 2025

  15. [23]

    Demystify mamba in vision: A linear attention perspective,

    D. Han, Z. Wang, Z. Xia, Y. Han, Y. Pu, C. Ge, J. Song, S. Song, B. Zheng, and G. Huang, “Demystify mamba in vision: A linear attention perspective,” Advances in Neural Information Processing Systems, vol. 37, pp. 127181–127203, 2024

  16. [24]

    Metaformer baselines for vision,

    W. Yu, C. Si, P . Zhou, M. Luo, Y. Zhou, J. Feng, S. Yan, and X. Wang, “Metaformer baselines for vision,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 46, no. 2, pp. 896–912, 2024

  17. [25]

    Rmt: Retentive networks meet vision transformers,

    Q. Fan, H. Huang, M. Chen, H. Liu, and R. He, “Rmt: Retentive networks meet vision transformers,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pp. 5641– 5651, 2024

  18. [26]

    Transnext: Robust foveal visual perception for vision transformers,

    D. Shi, “Transnext: Robust foveal visual perception for vision transformers,” in Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition, pp. 17773–17783, 2024

  19. [27]

    Segman: Omni-scale context modeling with state space models and local attention for semantic segmen- tation,

    Y. Fu, M. Lou, and Y. Yu, “Segman: Omni-scale context modeling with state space models and local attention for semantic segmen- tation,” in Proceedings of the Computer Vision and Pattern Recognition Conference, pp. 19077–19087, 2025

  20. [28]

    Unified perceptual parsing for scene understanding,

    T. Xiao, Y. Liu, B. Zhou, Y. Jiang, and J. Sun, “Unified perceptual parsing for scene understanding,” in European Conference on Com- puter Vision, pp. 418–434, 2018

  21. [29]

    Cascade r-cnn: High quality object detection and instance segmentation,

    Z. Cai and N. Vasconcelos, “Cascade r-cnn: High quality object detection and instance segmentation,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 43, no. 5, pp. 1483–1498, 2019

  22. [30]

    Moganet: Multi-order gated aggregation network,

    S. Li, Z. Wang, Z. Liu, C. Tan, H. Lin, D. Wu, Z. Chen, J. Zheng, and S. Z. Li, “Moganet: Multi-order gated aggregation network,” in International Conference on Learning Representations, 2023

  23. [31]

    Imagenet classifica- tion with deep convolutional neural networks,

    A. Krizhevsky, I. Sutskever, and G. E. Hinton, “Imagenet classifica- tion with deep convolutional neural networks,” Advances in neural information processing systems, vol. 25, 2012

  24. [32]

    Very deep convolutional net- works for large-scale image recognition,

    K. Simonyan and A. Zisserman, “Very deep convolutional net- works for large-scale image recognition,” in International Confer- ence on Learning Representations, 2015

  25. [33]

    Deep residual learning for image recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 770–778, 2016

  26. [34]

    Densely connected convolutional networks,

    G. Huang, Z. Liu, L. Van Der Maaten, and K. Q. Weinberger, “Densely connected convolutional networks,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition , pp. 4700–4708, 2017

  27. [35]

    Convnext v2: Co-designing and scaling convnets with masked autoencoders,

    S. Woo, S. Debnath, R. Hu, X. Chen, Z. Liu, I. S. Kweon, and S. Xie, “Convnext v2: Co-designing and scaling convnets with masked autoencoders,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 16133–16142, 2023

  28. [36]

    More convnets in the 2020s: Scaling up kernels beyond 51x51 using sparsity,

    S. Liu, T. Chen, X. Chen, X. Chen, Q. Xiao, B. Wu, M. Pechenizkiy, D. Mocanu, and Z. Wang, “More convnets in the 2020s: Scaling up kernels beyond 51x51 using sparsity,” inInternational Conference on Learning Representations, 2023

  29. [37]

    Unireplknet: A universal perception large-kernel convnet for audio, video, point cloud, time-series and image recognition,

    X. Ding, Y. Zhang, Y. Ge, S. Zhao, L. Song, X. Yue, and Y. Shan, “Unireplknet: A universal perception large-kernel convnet for audio, video, point cloud, time-series and image recognition,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024

  30. [38]

    Pelk: Parameter- efficient large kernel convnets with peripheral convolution,

    H. Chen, X. Chu, Y. Ren, X. Zhao, and K. Huang, “Pelk: Parameter- efficient large kernel convnets with peripheral convolution,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp. 5557–5567, 2024

  31. [39]

    Hor- net: Efficient high-order spatial interactions with recursive gated convolutions,

    Y. Rao, W. Zhao, Y. Tang, J. Zhou, S. N. Lim, and J. Lu, “Hor- net: Efficient high-order spatial interactions with recursive gated convolutions,” Advances in Neural Information Processing Systems , vol. 35, pp. 10353–10366, 2022

  32. [40]

    Focal modulation networks,

    J. Yang, C. Li, X. Dai, and J. Gao, “Focal modulation networks,” Advances in Neural Information Processing Systems, vol. 35, pp. 4203– 4217, 2022

  33. [41]

    Overlock: An overview-first-look-closely-next convnet with context-mixing dynamic kernels,

    M. Lou and Y. Yu, “Overlock: An overview-first-look-closely-next convnet with context-mixing dynamic kernels,” in Proceedings of the Computer Vision and Pattern Recognition Conference, pp. 128–138, 2025

  34. [42]

    Neural mech- anisms of visual attention: how top-down feedback highlights relevant locations,

    Y. B. Saalmann, I. N. Pigarev, and T. R. Vidyasagar, “Neural mech- anisms of visual attention: how top-down feedback highlights relevant locations,” Science, vol. 316, no. 5831, pp. 1612–1615, 2007

  35. [43]

    Cmt: Convolutional neural networks meet vision transformers,

    J. Guo, K. Han, H. Wu, Y. Tang, X. Chen, Y. Wang, and C. Xu, “Cmt: Convolutional neural networks meet vision transformers,” 14 in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 12175–12185, 2022

  36. [44]

    On the integration of self-attention and convolution,

    X. Pan, C. Ge, R. Lu, S. Song, G. Chen, Z. Huang, and G. Huang, “On the integration of self-attention and convolution,” in Proceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 815–825, 2022

  37. [45]

    Mixformer: Mixing features across windows and dimensions,

    Q. Chen, Q. Wu, J. Wang, Q. Hu, T. Hu, E. Ding, J. Cheng, and J. Wang, “Mixformer: Mixing features across windows and dimensions,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 5249–5259, 2022

  38. [46]

    Transxnet: Learning both global and local dynamics with a dual dynamic token mixer for visual recognition,

    M. Lou, S. Zhang, H.-Y. Zhou, S. Yang, C. Wu, and Y. Yu, “Transxnet: Learning both global and local dynamics with a dual dynamic token mixer for visual recognition,” IEEE Transactions on Neural Networks and Learning Systems , vol. 36, no. 6, pp. 11534– 11547, 2025

  39. [47]

    Uniformer: Unifying convolution and self-attention for visual recognition,

    K. Li, Y. Wang, J. Zhang, P . Gao, G. Song, Y. Liu, H. Li, and Y. Qiao, “Uniformer: Unifying convolution and self-attention for visual recognition,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 45, no. 10, pp. 12581–12600, 2023

  40. [48]

    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, pp. 633–641, 2017

  41. [49]

    The cityscapes dataset for semantic urban scene understanding,

    M. Cordts, M. Omran, S. Ramos, T. Rehfeld, M. Enzweiler, R. Benenson, U. Franke, S. Roth, and B. Schiele, “The cityscapes dataset for semantic urban scene understanding,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition , pp. 3213–3223, 2016

  42. [50]

    Microsoft coco: Common objects in context,

    T.-Y. Lin, M. Maire, S. Belongie, J. Hays, P . Perona, D. Ramanan, P . Doll´ar, and C. L. Zitnick, “Microsoft coco: Common objects in context,” in European Conference on Computer Vision , pp. 740–755, Springer, 2014

  43. [51]

    Segnext: Rethinking convolutional attention design for semantic segmentation,

    M.-H. Guo, C.-Z. Lu, Q. Hou, Z. Liu, M.-M. Cheng, and S.-M. Hu, “Segnext: Rethinking convolutional attention design for semantic segmentation,” Advances in neural information processing systems , vol. 35, pp. 1140–1156, 2022

  44. [52]

    Multi-scale representations by varying window attention for semantic segmentation,

    H. Yan, M. Wu, and C. Zhang, “Multi-scale representations by varying window attention for semantic segmentation,” in Interna- tional Conference on Learning Representations, 2024

  45. [53]

    A convnet for the 2020s,

    Z. Liu, H. Mao, C.-Y. Wu, C. Feichtenhofer, T. Darrell, and S. Xie, “A convnet for the 2020s,” in Proceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition , pp. 11976–11986, 2022

  46. [54]

    Deeplab: Semantic image segmentation with deep convo- lutional nets, atrous convolution, and fully connected crfs,

    L.-C. Chen, G. Papandreou, I. Kokkinos, K. Murphy, and A. L. Yuille, “Deeplab: Semantic image segmentation with deep convo- lutional nets, atrous convolution, and fully connected crfs,” IEEE Transactions on Pattern Analysis and Machine Intelligence , vol. 40, no. 4, pp. 834–848, 2017

  47. [55]

    Segformer: Simple and efficient design for semantic segmenta- tion with transformers,

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

  48. [56]

    Encoder-decoder with atrous separable convolution for semantic image segmentation,

    L.-C. Chen, Y. Zhu, G. Papandreou, F. Schroff, and H. Adam, “Encoder-decoder with atrous separable convolution for semantic image segmentation,” in Proceedings of the European conference on computer vision, pp. 801–818, 2018

  49. [57]

    Quadmamba: Learning quadtree-based selective scan for visual state space model,

    F. Xie, W. Zhang, Z. Wang, and C. Ma, “Quadmamba: Learning quadtree-based selective scan for visual state space model,” Ad- vances in Neural Information Processing Systems, vol. 37, pp. 117682– 117707, 2024

  50. [58]

    Inceptionnext: When incep- tion meets convnext,

    W. Yu, P . Zhou, S. Yan, and X. Wang, “Inceptionnext: When incep- tion meets convnext,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024

  51. [59]

    Multi-scale vmamba: Hierarchy in hierarchy visual state space model,

    Y. Shi, M. Dong, and C. Xu, “Multi-scale vmamba: Hierarchy in hierarchy visual state space model,” Advances in Neural Information Processing Systems, 2024

  52. [60]

    Densenets reloaded: Paradigm shift beyond resnets and vits,

    D. Kim, B. Heo, and D. Han, “Densenets reloaded: Paradigm shift beyond resnets and vits,” in European Conference on Computer Vision, 2024

  53. [61]

    Regionvit: Regional-to-local attention for vision transformers,

    C.-F. Chen, R. Panda, and Q. Fan, “Regionvit: Regional-to-local attention for vision transformers,” in International Conference on Learning Representations, 2022

  54. [62]

    Mpvit: Multi-path vision transformer for dense prediction,

    Y. Lee, J. Kim, J. Willette, and S. J. Hwang, “Mpvit: Multi-path vision transformer for dense prediction,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pp. 7287–7296, 2022

  55. [63]

    Conv2former: A simple transformer-style convnet for visual recognition,

    Q. Hou, C.-Z. Lu, M.-M. Cheng, and J. Feng, “Conv2former: A simple transformer-style convnet for visual recognition,” IEEE Transactions on Pattern Analysis and Machine Intelligence , vol. 46, no. 12, pp. 8274–8283, 2024

  56. [64]

    Scale-aware modula- tion meet transformer,

    W. Lin, Z. Wu, J. Chen, J. Huang, and L. Jin, “Scale-aware modula- tion meet transformer,” in Proceedings of the IEEE/CVF International Conference on Computer Vision, pp. 6015–6026, 2023

  57. [65]

    Internimage: Exploring large-scale vision foun- dation models with deformable convolutions,

    W. Wang, J. Dai, Z. Chen, Z. Huang, Z. Li, X. Zhu, X. Hu, T. Lu, L. Lu, H. Li, et al., “Internimage: Exploring large-scale vision foun- dation models with deformable convolutions,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2023

  58. [66]

    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,” in IEEE Conference on Computer Vision and Pattern Recognition, pp. 248–255, 2009

  59. [67]

    Decoupled weight decay regulariza- tion,

    I. Loshchilov and F. Hutter, “Decoupled weight decay regulariza- tion,” arXiv preprint arXiv:1711.05101, 2017

  60. [68]

    Deep networks with stochastic depth,

    G. Huang, Y. Sun, Z. Liu, D. Sedra, and K. Q. Weinberger, “Deep networks with stochastic depth,” in European Conference on Com- puter Vision, pp. 646–661, Springer, 2016

  61. [69]

    Mask r-cnn,

    K. He, G. Gkioxari, P . Doll ´ar, and R. Girshick, “Mask r-cnn,” in Proceedings of the IEEE International Conference on Computer Vision , pp. 2961–2969, 2017

  62. [70]

    Embedding-free transformer with inference spatial reduction for efficient semantic segmentation,

    H. Yu, Y. Cho, B. Kang, S. Moon, K. Kong, and S.-J. Kang, “Embedding-free transformer with inference spatial reduction for efficient semantic segmentation,” in European Conference on Com- puter Vision, pp. 92–110, Springer, 2024

  63. [71]

    Context-guided spatial feature reconstruction for efficient semantic segmentation,

    Z. Ni, X. Chen, Y. Zhai, Y. Tang, and Y. Wang, “Context-guided spatial feature reconstruction for efficient semantic segmentation,” in European Conference on Computer Vision , pp. 239–255, Springer, 2024

  64. [72]

    Vit-comer: Vision transformer with convolutional multi-scale feature interaction for dense predictions,

    C. Xia, X. Wang, F. Lv, X. Hao, and Y. Shi, “Vit-comer: Vision transformer with convolutional multi-scale feature interaction for dense predictions,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp. 5493–5502, 2024

  65. [73]

    Object-contextual representations for semantic segmentation,

    Y. Yuan, X. Chen, and J. Wang, “Object-contextual representations for semantic segmentation,” in European Conference on Computer Vision, pp. 173–190, Springer, 2020

  66. [74]

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

    B. Cheng, A. Schwing, and A. Kirillov, “Per-pixel classification is not all you need for semantic segmentation,” Advances in neural information processing systems, vol. 34, pp. 17864–17875, 2021

  67. [75]

    Masked-attention mask transformer for universal image segmen- tation,

    B. Cheng, I. Misra, A. G. Schwing, A. Kirillov, and R. Girdhar, “Masked-attention mask transformer for universal image segmen- tation,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp. 1290–1299, 2022

  68. [76]

    Feedformer: Revisit- ing transformer decoder for efficient semantic segmentation,

    J.-h. Shim, H. Yu, K. Kong, and S.-J. Kang, “Feedformer: Revisit- ing transformer decoder for efficient semantic segmentation,” in Proceedings of the AAAI conference on artificial intelligence , vol. 37, pp. 2263–2271, 2023

  69. [77]

    Low-resolution self-attention for semantic segmentation,

    Y.-H. Wu, S.-C. Zhang, Y. Liu, L. Zhang, X. Zhan, D. Zhou, J. Feng, M.-M. Cheng, and L. Zhen, “Low-resolution self-attention for semantic segmentation,” IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025

  70. [78]

    Understanding the effective receptive field in deep convolutional neural networks,

    W. Luo, Y. Li, R. Urtasun, and R. Zemel, “Understanding the effective receptive field in deep convolutional neural networks,” Advances in Neural Information Processing Systems , vol. 29, 2016

  71. [79]

    Deformable convnets v2: More deformable, better results,

    X. Zhu, H. Hu, S. Lin, and J. Dai, “Deformable convnets v2: More deformable, better results,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pp. 9308– 9316, 2019

Pith tools

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