Pith. sign in

REVIEW 4 major objections 6 minor 48 references

InceptionMamba: An Efficient Hybrid Network with Large Band Convolution and Bottleneck Mamba

T0 review · 4 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read InceptionMamba claims state-of-the-art image classification by pairing orthogonal band convolutions with a bottleneck Mamba module, reaching 84.7% top-1 on ImageNet-1K at 83M parameters and 14.3G FLOPs.

desk verdict A plausible incremental hybrid backbone whose SOTA claim is undercut by unretrained baseline comparisons; worth reviewing, not yet proven. read the letter →

arxiv 2506.08735 v3 pith:K3QTAAXY submitted 2025-06-10 cs.CV

Yuhang Wang , Jun Li , Zhijian Wu , Jifeng Shen , Jianhua Xu , Wankou Yang This is my paper
classification cs.CV MSC 68T0768T45
keywords InceptionMambaOrthogonalBandConvolutionsGlobalContextualModelingBottleneckCross-channelInformationFusionVisualStateSpaceModelsImageNetclassification
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

This paper proposes InceptionMamba, a vision backbone that upgrades InceptionNeXt by replacing its parallel one-dimensional strip convolutions with orthogonal band convolutions and adding a bottleneck Mamba module for global context. The aim is to show that a CNN-style efficient parallel structure can gain long-range modeling without the quadratic cost of attention or the overhead of a full Mamba token mixer. On ImageNet-1K the three variants report 83.1%, 84.1%, and 84.7% top-1 accuracy at 25M/46M/83M parameters and 4.0G/7.8G/14.3G FLOPs, beating listed CNN and Mamba baselines at comparable or lower compute. The same backbone also improves COCO detection and ADE20K segmentation under standard detectors, so the claim is that this design is a generally stronger efficient backbone, not just a classification trick.

What carries the argument

The core object is the InceptionMamba block, made of a ConvMixer and a GlobalMixer. ConvMixer is the multi-branch depthwise module in which orthogonal band convolutions (3x11 and 11x3) act on a fraction of channels to capture both horizontal and vertical cues; GlobalMixer is a bottleneck around the SS2D module, a selective-scan state-space operator that scans the feature map in multiple directions, enabling long-range dependencies with linear complexity. The bottleneck (channel compression ratio 2) is what keeps the SS2D cheap: the ablation reports the same 83.1% top-1 as unfettered SS2D while cutting parameters from 29M to 25M and FLOPs from 4.6G to 4.0G.

What would settle it

Retrain InceptionNeXt-T and VMamba-T from scratch with the exact 300-epoch AdamW pipeline described in Section 4.1 (batch 512, RandAugment, Mixup, label smoothing 0.1); if either reaches or exceeds 83.1% top-1 at 224x224, the claimed accuracy advantage over the paper's central baselines collapses.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that InceptionNeXt's two limitations—weak response to orthogonal structures from strip convolutions, and no global receptive field—can be fixed together without efficiency loss. The ConvMixer splits channels into three groups and applies 3x3 depthwise convolution to one group, orthogonal 3x11 and 11x3 band convolutions to a second, and identity to the largest group, giving cohesive local spatial modeling. The GlobalMixer then compresses channels by half with a 1x1 convolution, runs the compressed features through the SS2D selective-scan state-space module, and expands back, which the paper argues fuses cross-channel information and enlarges the receptive field at linear complexity. The reported result is consistent gains across model scales and tasks: 83.1/84.1/84.7% ImageNet top-1, 46.0/47.5/48.1 box AP on COCO with Mask R-CNN, and 47.3/49.2/50.1 mIoU on ADE20K, with fewer parameters and FLOPs than most compared backbones.

Load-bearing premise

The paper takes every competing model's published accuracy as a fair baseline without retraining those models under InceptionMamba's exact recipe, so the reported edge could be partly training-schedule rather than architecture.

Editorial extensions

If this is right

  • At the tiny scale, InceptionMamba-T reports 83.1% top-1 with 25M parameters and 4.0G FLOPs, ahead of VMamba-T (82.2%) and InceptionNeXt-T (82.3%) at similar or lower compute.
  • The bottleneck Mamba is the efficiency lever: replacing full SS2D with bottleneck SS2D keeps accuracy at 83.1% while reducing parameters by 13.8% and FLOPs by 15.2%, and raises throughput from 1042 to 1362 images/s.
  • The gains carry over to dense prediction, with InceptionMamba-B reaching 48.1% box AP on COCO and 50.1% mIoU on ADE20K, topping the compared baselines at lower FLOPs.
  • If the comparisons hold, architectural families do not need to choose between local efficiency and global context: a small channel bottleneck before a selective scan is enough to add long-range modeling.

Reading between the lines

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

  • A testable extension is to drop or shrink the identity branch's 75% channel share to see how much of the gain is due to the orthogonal band branch alone; the paper's branch-ratio ablation suggests sensitivity is low, so the band convolutions may matter more than the exact allocation.
  • The bottleneck-before-scan design implies that global context in this hybrid does not need full channel width, a principle that could be carried to other Mamba-based backbones to cut cost without retraining the whole pipeline.
  • The CAM visualizations are suggestive but not quantified; measuring the receptive-field radius or overlap between high-activation regions would give a direct test of the 'cohesive spatial modeling' claim beyond top-1 accuracy.
  • Because baseline numbers are quoted from prior papers, an independent fair-recipe retraining of all compared models is the cleanest way to confirm the claimed margin; until then the 0.3–0.9 point gaps should be read as upper bounds.
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

4 major / 6 minor

Summary. The paper proposes InceptionMamba, a hybrid CNN-Mamba backbone that modifies InceptionNeXt by replacing one-dimensional strip convolutions with orthogonal band convolutions (3x11 and 11x3) and inserting a GlobalMixer that applies a bottleneck structure with an SS2D state-space module for cross-channel fusion and global context modeling. The architecture is evaluated on ImageNet-1K classification, COCO object detection and instance segmentation, and ADE20K semantic segmentation. The authors report state-of-the-art results, e.g., InceptionMamba-B attaining 84.7% top-1 accuracy on ImageNet-1K with 83M parameters and 14.3G FLOPs, and support the design with ablation studies over the branch ratio, ConvMixer structures, and the bottleneck/SS2D module.

Significance. If the reported numbers are reproducible, InceptionMamba contributes a useful design point in efficient hybrid backbones, combining the local modeling strengths of convolutional networks with the long-range modeling of state-space models while maintaining linear complexity. The paper covers three standard benchmarks and includes multiple ablations, which is a solid empirical package. The CAM visualizations in Figures 2 and 6 are informative. However, the central claim of state-of-the-art accuracy rests on cross-paper comparisons without controlled retraining of baselines, and there are internal inconsistencies in parameter counts and in the formal definition of the GlobalMixer. These issues must be resolved before the claims can be accepted.

major comments (4)
  1. [Section 4.1, Table 2] The central state-of-the-art claim rests on comparisons against published numbers from prior papers without retraining those baselines under the training protocol used for InceptionMamba. The decisive margins over RDNet-B (84.7 vs 84.4) and MambaOut-B (84.7 vs 84.2) are 0.3-0.5 points, exactly the range in which differences in augmentation, stochastic depth, label smoothing, and schedule can shift from-scratch ImageNet results. The paper reports a single run with no variance estimate, and the setup section lists both RandAugment and AutoAugment, making the recipe ambiguous. Without a control experiment that retrains the nearest competitors under identical conditions, the claimed superiority is not established.
  2. [Section 3.3, Eq. (12)] Equation (12) defines the GlobalMixer output as Y = X + X', where X is the ConvMixer input and X' is the ConvMixer output, but the intermediate lines apply SS2D and two 1x1 convolutions to produce X''; the final Y should be the residual sum involving the transformed X'', not X'. Additionally, the variable X'' is overwritten three times. This formal description does not match the architecture in Fig. 4 and prevents accurate reimplementation. Please rewrite Eq. (12) with distinct variables and a correct residual term.
  3. [Tables 1, 5, 6, 7, 8 and Section 3.3] The reported parameter counts are inconsistent across the paper. Table 1 lists InceptionMamba-T at 25M, Table 7 reports 26M for the same configuration, and the text in Section 3.3 claims a reduction from 29M to 25M (13.8%), while Table 7 shows 29M to 26M (10.3%). Table 8 reports a 21.4M variant without the bottleneck and a 25.4M model with it, while Table 6 lists 25.4M for all ConvMixer variants. These discrepancies make it impossible to verify the efficiency claims and should be reconciled.
  4. [Section 4.4, Tables 5-8] The hyperparameters of the architecture, including the branch ratio [0.125,0.125,0.75], the band kernel sizes 3x11 and 11x3, and the bottleneck ratio 1/2, are selected via ablations on ImageNet-1K, and the final accuracies are then reported on the same dataset. Because the same validation split is used for both model selection and final evaluation, the headline numbers may be optimistically biased. A held-out split or an explicit statement of the validation protocol used for design choices is needed to support the state-of-the-art claim.
minor comments (6)
  1. [Table 2 and Reference [36]] The model name is spelled 'QuardMamba' in Table 2 but 'QuadMamba' in reference [36]; please align the spelling.
  2. [Tables 7 and 8] The abbreviation 'TP.' used for throughput is not defined in the text or table captions; please spell out the unit or add a full definition.
  3. [Section 4.1, experimental setup] The setup text says the training incorporates RandAugment and then states that AutoAugment is adopted following [31]; please clarify which augmentation policy is actually used, as this ambiguity is important for reproducibility.
  4. [Figure 1 caption] The caption reads 'at 2242 resolutions' in the manuscript text; this should be 'at 224x224 resolution'.
  5. [References [19] and [20]] Reference [20] appears to be a duplicate of [19] with the same title and venue; the Inception-v2/v3 citation should be distinguished, for example by citing the correct source for batch normalization.
  6. [General] The paper states that source code will be available, but the repository is not provided at submission time; given the need for controlled re-benchmarking described above, making the code and model weights available upon revision would substantially strengthen verification.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: InceptionMamba's SOTA claim rests on external benchmarks and explicit architecture equations; ablations are standard hyperparameter selection, not circular, and no load-bearing self-citations appear.

full rationale

The paper's derivation chain is empirical and self-contained against external benchmarks. The architecture is defined explicitly: ConvMixer splits channels and applies 3x3, 3x11, 11x3 depthwise convolutions (Eq. 10), and GlobalMixer applies channel compression, SS2D, and expansion (Eq. 12). The contributions are then tested by ablations (Tables 5-8) that compare the proposed modules to alternatives (strip convolution, plain DWConv, attention, GELU, etc.) under matched parameter/FLOP budgets. Selecting the branch ratio [0.125, 0.125, 0.75] and bottleneck ratio from these ablations is ordinary hyperparameter selection on an external benchmark, not fitting a parameter to a subset and then renaming the fit as a prediction. No load-bearing claim is justified by self-citation: all references to InceptionNeXt, SLaK, Mamba, VMamba, etc. are prior external work by non-overlapping authors, and no uniqueness theorem or ansatz is imported from the present authors' earlier work. 'Band convolution' is explicitly defined in Eq. (10) as 3x11 and 11x3 depthwise convolutions, so it is not a renamed known result. The main weakness, cross-paper baseline numbers in Table 2 without same-recipe retraining or variance estimates, is a correctness/reproducibility risk, not circularity. Accordingly, no circular step can be exhibited and the score is 0.

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

The architecture relies on several domain assumptions inherited from prior work, and its performance numbers are generated from empirical training rather than a derivation. The main free parameters (branch ratio, kernel size) were chosen on the test benchmark, which is common in this literature but should temper claims of prediction. No new physical or conceptual entities are introduced.

free parameters (3)
  • Branch ratio [0.125, 0.125, 0.75] = 0.125 square, 0.125 band, 0.75 identity
    Chosen empirically via ablation on ImageNet (Table 5).
  • Band kernel size 3x11 and 11x3 = (3,11)
    Hand-picked; no search over kernel sizes is reported.
  • Bottleneck ratio 1/2 = 0.5
    Fixed in all blocks; the ratio is not varied in any ablation.
assumptions (3)
  • domain assumption SS2D (bidirectional cross-scan Mamba layer) provides effective global context modeling with linear complexity.
    Borrowed from VMamba; the paper does not independently verify this for its architecture beyond comparing Bottleneck+SS2D to Bottleneck+attention.
  • domain assumption The reported baseline accuracies from prior publications are accurate and comparable to numbers produced under InceptionMamba's training setup.
    The paper uses published numbers without retraining; differences in augmentation or schedule could shift rankings.
  • domain assumption The standard training recipe (AdamW, 300 epochs, RandAugment, Mixup) is appropriate for all compared architectures.
    The paper follows the recipe of [31] but does not tune it per baseline.

how reviews work

0 comments
Cite this review

Pith. "Pith review of InceptionMamba: An Efficient Hybrid Network with Large Band Convolution and Bottleneck Mamba." pith.science (2026). https://pith.science/paper/K3QTAAXY

@misc{pith2026250608735,
  author       = {Pith},
  title        = {Pith review of: InceptionMamba: An Efficient Hybrid Network with Large Band Convolution and Bottleneck Mamba},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/K3QTAAXY}},
  note         = {Machine review of arXiv:2506.08735}
}
read the original abstract

Within the family of convolutional neural networks, InceptionNeXt has shown excellent competitiveness in image classification and a number of downstream tasks. Built on parallel one-dimensional strip convolutions, however, it suffers from limited ability of capturing spatial dependencies along different dimensions and fails to fully explore spatial modeling in local neighborhood. Besides, inherent locality constraints of convolution operations are detrimental to effective global context modeling. To overcome these limitations, we propose a novel backbone architecture termed InceptionMamba in this study. More specifically, the traditional one-dimensional strip convolutions are replaced by orthogonal band convolutions in our InceptionMamba to achieve cohesive spatial modeling. Furthermore, global contextual modeling can be achieved via a bottleneck Mamba module, facilitating enhanced cross-channel information fusion and enlarged receptive field. Extensive evaluations on classification and various downstream tasks demonstrate that the proposed InceptionMamba achieves state-of-the-art performance with superior parameter and computational efficiency. The source code will be available at https://github.com/Wake1021/InceptionMamba.

Figures

Figures reproduced from arXiv: 2506.08735 by the authors.

Figure 1
Figure 1. Performance on ImageNet-1K validation set at 224 [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Visualizations of different CAMs [15] generated from ConvNeXt(b), InceptionNeXt(c) and our InceptionMamba(d) given a panda image(a). It can be observed that InceptionMamba showcases superior spatial modeling capacity in capturing semantic-aware areas in the neighborhood of panda eye and body. In contrast, InceptionNeXt only focuses on scattered eye-centric regions, yielding less cohesive representations. Mamba terme… view at source ↗
Figure 3
Figure 3. Our InceptionMamba architecture with four stages. Similar to ConvNeXt [3], InceptionMamba employs hierarchical architecture of four consecutive stages. Each stage consists of a patch embedding layer or a downsampling module, combined with Ni InceptionMamba blocks. Recent advancements in visual Mamba models demonstrate the potential of SSM in a variety of vision tasks. In Vision Mamba [25], a bidirectional Mamba bloc… view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Illustrative comparison of ConvNeXt, InceptionNeXt, and our InceptionMamba block. Inception [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]
Figure 5
Figure 5. Figure 5: Visualized feature maps along consecutive channels in an intermediate layer of a pretrained [PITH_FULL_IMAGE:figures/full_fig_p012_5.png]
Figure 6
Figure 6. Figure 6: Comparison of CAMs generated from different mainstream architectures. Our InceptionMamba is capable of characterizing key semantic-aware regions, which is beneficial for a variety of downstream vision tasks. itself as a potential state-of-the-art lightweight architectu…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

48 extracted references · 46 canonical work pages

  1. [1]

    LeCun, B

    Y . LeCun, B. Boser, J. S. Denker, D. Henderson, R. E. Howard, W. Hubbard, L. D. Jackel, Backpropagation applied to handwritten zip code recognition, Neu- ral Computation 1 (4) (1989) 541–551

    work page 1989

  2. [2]

    LeCun, L

    Y . LeCun, L. Bottou, Y . Bengio, P. Haffner, Gradient-based learning applied to document recognition, Proceedings of the IEEE 86 (11) (1998) 2278–2324

    work page 1998

  3. [3]

    Z. Liu, H. Mao, C.-Y . Wu, C. Feichtenhofer, T. Darrell, S. Xie, A ConvNet for the 2020s, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022, pp. 11966–11976

    work page 2022

  4. [4]

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

    work page 2016

  5. [5]

    Huang, Z

    G. Huang, Z. Liu, L. van der Maaten, K. Q. Weinberger, Densely connected con- volutional networks, in: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017, pp. 2261–2269

    work page 2017

  6. [6]

    W. Yu, P. Zhou, S. Yan, X. Wang, Inceptionnext: When inception meets convnext, in: Proceedings of the IEEE /CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024, pp. 5672–5683

    work page 2024

  7. [7]

    X. Ding, X. Zhang, J. Han, G. Ding, Scaling up your kernels to 31×31: Revisiting large kernel design in CNNs, in: Proceedings of the IEEE /CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022, pp. 11953–11965. 23

    work page 2022

  8. [8]

    F. Yu, V . Koltun, Multi-scale context aggregation by dilated convolutions, in: Pro- ceedings of the International Conference on Learning Representations (ICLR), 2016, pp. 1–13

    work page 2016

Show all 48 references
  1. [9]

    Dosovitskiy, L

    A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gelly, J. Uszkoreit, N. Houlsby, An image is worth 16x16 words: Transformers for image recognition at scale, in: Proceedings of International Conference o...

    2021

  2. [10]

    Tolstikhin, N

    I. Tolstikhin, N. Houlsby, A. Kolesnikov, L. Beyer, X. Zhai, T. Unterthiner, J. Yung, A. Steiner, D. Keysers, J. Uszkoreit, M. Lucic, A. Dosovitskiy, MLP- Mixer: An all-MLP architecture for vision, in: Advances in Neural Information Processing Systems (NeurIPS), 2021, pp. 1–12

    2021

  3. [11]

    Steiner, A

    A. Steiner, A. Kolesnikov, , X. Zhai, R. Wightman, J. Uszkoreit, L. Beyer, How to train your ViT? data, augmentation, and regularization in Vision Transformers, arXiv preprint arXiv:2106.10270 (2021)

    2021 arXiv

  4. [12]

    Chen, C.-J

    X. Chen, C.-J. Hsieh, B. Gong, When Vision Transformers outperform ResNets without pretraining or strong data augmentations, in: Proceedings of International Conference on Learning Representations (ICLR), 2022, pp. 1–20

    2022

  5. [13]

    Zhuang, B

    J. Zhuang, B. Gong, L. Yuan, Y . Cui, H. Adam, N. Dvornek, S. Tatikonda, J. Dun- can, T. Liu, Surrogate gap minimization improves sharpness-aware training, in: Proceedings of International Conference on Learning Representations (ICLR), 2022, pp. 1–24

    2022

  6. [14]

    X. Zhai, X. Wang, B. Mustafa, A. Steiner, D. Keysers, A. Kolesnikov, L. Beyer, LiT: Zero-shot transfer with locked-image text tuning, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022, pp. 18102–18112

    2022

  7. [15]

    R. R. Selvaraju, M. Cogswell, A. Das, R. Vedantam, D. Parikh, D. Batra, Grad-CAM: Visual explanations from deep networks via gradient-based localiza- 24 tion, in: Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2017, pp. 618–626

    2017

  8. [16]

    Hatamizadeh, J

    A. Hatamizadeh, J. Kautz, Mambavision: A hybrid mamba-transformer vision backbone, in: Proceedings of the IEEE /CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025, pp. 25261–25270

    2025

  9. [17]

    Y . Liu, Y . Tian, Y . Zhao, H. Yu, L. Xie, Y . Wang, Q. Ye, Y . Liu, VMamba: Visual state space model, in: Advances in Neural Information Processing Systems (NeurIPS), 2024, pp. 103031–103063

    2024

  10. [18]

    Szegedy, W

    C. Szegedy, W. Liu, Y . Jia, P. Sermanet, S. Reed, D. Anguelov, D. Erhan, V . Van- houcke, A. Rabinovich, Going deeper with convolutions, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2015, pp. 1–9

    2015

  11. [19]

    Szegedy, V

    C. Szegedy, V . Vanhoucke, S. Ioffe, J. Shlens, Z. Wojna, Rethinking the inception architecture for computer vision, in: Proceedings of the IEEE /CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2016, pp. 2818–2826

    2016

  12. [20]

    Szegedy, V

    C. Szegedy, V . Vanhoucke, S. Ioffe, J. Shlens, Z. Wojna, Rethinking the inception architecture for computer vision, in: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016, pp. 2818–2826

    2016

  13. [21]

    Szegedy, S

    C. Szegedy, S. Io ffe, V . Vanhoucke, A. Alemi, Inception-v4, inception-resnet and the impact of residual connections on learning, in: Proceedings of the AAAI Conference on Artificial Intelligence, 2017, pp. 4278–4284

    2017

  14. [22]

    X. Ding, X. Zhang, J. Han, G. Ding, Diverse branch block: Building a convolu- tion as an inception-like unit, in: Proceedings of the IEEE /CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021, pp. 10881–10890

    2021

  15. [23]

    A. Gu, T. Dao, Mamba: Linear-time sequence modeling with selective state spaces, arXiv preprint arXiv:2312.00752 (2023). 25

    2023 arXiv

  16. [24]

    T. Dao, A. Gu, Transformers are SSMs: Generalized models and e fficient al- gorithms through structured state space duality, in: International Conference on Machine Learning (ICML), 2024, pp. 10041–10071

    2024

  17. [25]

    L. Zhu, B. Liao, Q. Zhang, X. Wang, W. Liu, X. Wang, Vision Mamba: E ffi- cient visual representation learning with bidirectional state space model, in: Pro- ceedings of International Conference on Machine Learning (ICML), 2024, pp. 62429–62442

    2024

  18. [26]

    W. Yu, X. Wang, MambaOut: Do we really need Mamba for vision?, in: Proceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025, pp. 4484–4496

    2025

  19. [27]

    S. Liu, T. Chen, X. Chen, X. Chen, Q. Xiao, B. Wu, M. Pechenizkiy, D. Mocanu, Z. Wang, More convnets in the 2020s: Scaling up kernels beyond 51x51 using sparsity, in: Proceedings of International Conference on Learning Representa- tions (ICLR), 2022, pp. 1–23

    2022

  20. [28]

    J. Deng, W. Dong, R. Socher, L. Li, K. Li, L. Fei-Fei, Imagenet: A large-scale hierarchical image database, in: Proceedings of the IEEE /CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2009, pp. 248–255

    2009

  21. [29]

    Russakovsky, J

    O. Russakovsky, J. Deng, H. Su, J. Krause, S. Satheesh, S. Ma, Z. Huang, A. Karpathy, A. Khosla, M. Bernstein, et al., Imagenet large scale visual recogni- tion challenge, International Journal of Computer Vision 115 (3) (2015) 211–252

    2015

  22. [30]

    Loshchilov, F

    I. Loshchilov, F. Hutter, Decoupled weight decay regularization, in: Proceedings of International Conference on Learning Representations (ICLR), 2019, pp. 1–22

    2019

  23. [31]

    H. He, J. Zhang, Y . Cai, H. Chen, X. Hu, Z. Gan, Y . Wang, C. Wang, Y . Wu, L. Xie, MobileMamba: Lightweight multi-receptive visual Mamba network, Pro- ceedings of the IEEE /CVF Conference on Computer Vision and Pattern Recog- nition (CVPR) (2025)

    2025

  24. [32]

    Y . Duan, W. Wang, Z. Chen, X. Zhu, L. Lu, T. Lu, Y . Qiao, H. Li, J. Dai, W. Wang, Vision-RWKV: Efficient and scalable visual perception with rwkv-like architec- 26 tures, in: Proceedings of International Conference on Learning Representations (ICLR), 2025, pp. 1–17

    2025

  25. [33]

    Alkin, M

    B. Alkin, M. Beck, K. P ¨oppel, S. Hochreiter, J. Brandstetter, Vision-LSTM: xL- STM as generic vision backbone, in: Proceedings of International Conference on Learning Representations (ICLR), 2025, pp. 1–22

    2025

  26. [34]

    C. Yang, Z. Chen, M. Espinosa, L. Ericsson, Z. Wang, J. Liu, E. J. Crowley, PlainMamba: Improving non-hierarchical Mamba in visual recognition, in: Pro- ceedings of British Machine Vision Conferenc (BMVC), 2024, pp. 1–22

    2024

  27. [35]

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

    2025

  28. [36]

    F. Xie, W. Zhang, Z. Wang, C. Ma, QuadMamba: Learning quadtree-based se- lective scan for visual state space model, in: Advances in Neural Information Processing Systems (NeurIPS), 2024, pp. 117682–117707

    2024

  29. [37]

    H. Chen, X. Chu, Y . Ren, X. Zhao, K. Huang, PeLK: Parameter-e fficient large kernel convnets with peripheral convolution, in: Proceedings of the IEEE /CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024, pp. 5557–5567

    2024

  30. [38]

    D. Kim, B. Heo, D. Han, DenseNets Reloaded: Paradigm shift beyond ResNets and ViTs, in: Proceedings of European Conference on Computer Vision (ECCV), 2024, pp. 395–415

    2024

  31. [39]

    T.-Y . Lin, M. Maire, S. Belongie, J. Hays, P. Perona, D. Ramanan, P. Doll ´ar, C. L. Zitnick, Microsoft coco: Common objects in context, in: Proceedings of the European Conference on Computer Vision (ECCV), 2014, pp. 740–755

    2014

  32. [40]

    K. He, G. Gkioxari, P. Doll ´ar, R. Girshick, Mask R-CNN, in: Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2017, pp. 2961– 2969. 27

    2017

  33. [41]

    Z. Liu, Y . Lin, Y . Cao, H. Hu, Y . Wei, Z. Zhang, S. Lin, B. Guo, Swin Trans- former: Hierarchical Vision Transformer using shifted windows, in: Proceedings of the IEEE /CVF International Conference on Computer Vision (ICCV), 2021, pp. 9992–10002

    2021

  34. [42]

    Z. Chen, Y . Duan, W. Wang, J. He, T. Lu, J. Dai, Y . Qiao, Vision transformer adapter for dense predictions, in: Proceedings of the International Conference on Learning Representations (ICLR), 2022, pp. 1–20

    2022

  35. [43]

    W. Wang, E. Xie, X. Li, D.-P. Fan, K. Song, D. Liang, T. Lu, P. Luo, L. Shao, Pvtv2: Improved baselines with pyramid vision transformer, Computational Vi- sual Media 8 (3) (2022) 1–10

    2022

  36. [44]

    J. Yang, C. Li, X. Dai, J. Gao, Focal modulation networks, in: Advances in Neural Information Processing Systems (NeurIPS), 2022, pp. 4203–4217

    2022

  37. [45]

    B. Zhou, H. Zhao, X. Puig, S. Fidler, A. Barriuso, A. Torralba, Scene parsing through ade20k dataset, in: Proceedings of the IEEE /CVF Conference on Com- puter Vision and Pattern Recognition (CVPR), 2017, pp. 633–641

    2017

  38. [46]

    T. Xiao, Y . Liu, B. Zhou, Y . Jiang, J. Sun, Unified perceptual parsing for scene understanding, in: Proceedings of the European Conference on Computer Vision (ECCV), 2018, pp. 418–434

    2018

  39. [47]

    Contributors, MMSegmentation: Openmmlab semantic segmentation toolbox and benchmark (2020)

    M. Contributors, MMSegmentation: Openmmlab semantic segmentation toolbox and benchmark (2020)

    2020

  40. [48]

    H. Cai, J. Li, M. Hu, C. Gan, S. Han, E fficientvit: Lightweight multi-scale atten- tion for high-resolution dense prediction, in: Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition (CVPR), 2023, pp. 17302– 17313. 28

    2023

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.