Pith. sign in

REVIEW 5 major objections 5 minor 84 references

A ring-based scan order for vision state-space models claims ImageNet 84.5% top-1 at 3.9G FLOPs while staying stable under rotation.

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

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

T0 review · deepseek-v4-flash

2026-08-03 04:41 UTC pith:UIYKYCV4

load-bearing objection Novel ring scan and strong ImageNet/COCO numbers, but the rotation-robustness mechanism in Sec. 3.3 is not correct as written and code is missing. the 5 major comments →

arxiv 2602.04170 v2 pith:UIYKYCV4 submitted 2026-02-04 cs.CV

Partial Ring Scan: Revisiting Scan Order in Vision State Space Models

Yi-Kuan Hsieh , Kuan-Chuan Peng , Xin li , Ming-Ching Chang , Yu-Chee Tseng , Jun-Wei Hsieh This is my paper
classification cs.CV MSC 68T4568T07
keywords state space modelsvision SSMsscan orderrotation robustnessring scanpartial channel filteringimage classificationobject detection
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper argues that the way a vision state-space model serializes a 2D image into a 1D token sequence is not a neutral implementation detail: it determines which pixels are adjacent to the recurrent state, and fixed raster-like paths fracture object continuity and degrade under rotation. To fix this, PRISMamba partitions the image into concentric rings, runs a short recurrent loop around each ring, averages the loop outputs into a ring descriptor, and propagates descriptors from inner to outer rings through a short radial SSM. Because rotation only permutes positions within a ring instead of remapping the whole path, the scan becomes rotation-stable without any rotation-specific training. A partial channel filter routes only high-salience channels through the recurrent ring pathway and keeps the rest on a residual branch, cutting FLOPs and increasing throughput while slightly improving accuracy. On ImageNet-1K the paper reports 84.5% top-1 at 3.9G FLOPs and 3,054 img/s, outperforming fixed-path baselines and staying near-flat under 30° and 60° rotations.

Core claim

The central claim is that ring-by-ring order-agnostic aggregation plus radial composition resolves the sequence–geometry mismatch in path-based vision SSMs. Pixels are assigned to rings by their Euclidean distance from the image center; within each ring an alternating clockwise/counterclockwise loop runs a selective SSM and the per-step outputs are averaged to a ring descriptor, making the within-ring representation independent of the starting point. A second SSM reads the short sequence of ring descriptors from innermost to outermost, and the final ring states are projected back onto the grid and fused residually. Partial channel filtering selects channels by a global-average-pooling thresh

What carries the argument

The key object is the ring partition defined by radius bins around the image center, combined with an alternating loop traversal per ring and order-agnostic aggregation (the average of per-step SSM outputs) to form a ring descriptor. This converts a global, path-dependent serialization into a short radial chain of ring descriptors, so rotation becomes a cyclic shift inside each ring rather than a global reindexing. The second mechanism, partial channel filtering, thresholds per-channel global-average activations and routes only salient channels through the recurrent ring pathway, with the remainder going through an identity residual branch; this is a hard, linear-cost channel routing that th

Load-bearing premise

The rotation-stability explanation assumes that rotating an image only cyclically shifts the tokens inside each ring; with rendered rotations that introduce padding and interpolation, that assumption does not strictly hold.

What would settle it

Take one ImageNet validation image, rotate it by 30° with standard interpolation and padding, apply the ring partition at the same center, and compare the set of token coordinates assigned to each ring before and after rotation. If any ring's membership changes by more than a cyclic shift (e.g., pixels enter or leave the ring), the paper's stated mechanism for rotation robustness is violated, and the flat rotation accuracy must be explained by some other property of ring aggregation.

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

If this is right

  • If ring scan is as rotation-stable as reported, rotation augmentation becomes nearly free for vision SSMs, and models no longer need flip-only augmentation to preserve scan continuity.
  • Treating scan order as a first-class design choice opens a new axis for accuracy–efficiency tuning in SSM backbones, alongside channel routing.
  • The reported ImageNet and COCO numbers imply that ring-style traversal can improve dense prediction (detection and segmentation) without increasing model size or training schedule.
  • Partial channel filtering, shown as a plug-in across several SSM families, suggests a general way to cut FLOPs and boost throughput in recurrent vision architectures.
  • Because token count per ring is small and the radial chain is short, the approach preserves linear-time complexity and stays within memory budgets of standard vision SSMs.

Where Pith is reading between the lines

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

  • The rotation-robustness mechanism as stated (rotation equals a cyclic shift within each ring) is exact only for rotations that map the discrete grid onto itself; for ordinary rendered rotations with interpolation and padding, ring membership can change at boundaries, so the flat rotation curves the paper reports may be driven partly by averaging and robustness of aggregated descriptors rather than
  • If an object detector supplies the ring center, the same traversal could become object-centric, potentially improving recognition of off-center subjects; the paper mentions this as future work, and it is a natural testable extension.
  • A direct stress test would be to compare PRISMamba with and without rotation augmentation, or to feed exact circularly rotated images instead of canvas rotations, to separate the cyclic-shift mechanism from interpolation artifacts.
  • The mean-threshold channel filter uses aggregate activation magnitude; a distribution-aware threshold might be more stable for small feature maps or unusual channel statistics, but the paper does not explore this.

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

5 major / 5 minor

Summary. The paper argues that scan order is a first-class design choice in Vision State Space Models and that conventional fixed-path scans (raster, serpentine, etc.) are brittle under geometric transformations such as in-plane rotation. It proposes PRISMamba, which partitions the image into concentric rings, runs an order-agnostic selective SSM over each ring, propagates features radially inward-to-outward with a short SSM, and uses a partial channel filtering mechanism to route only salient channels through the recurrent path. The claimed results are: 84.5% ImageNet-1K Top-1 at 3.9G FLOPs and 3054 img/s on an A100, COCO detection/segmentation gains over VMamba and GroupMamba at lower FLOPs, and near-constant accuracy under 30/60-degree rotation while fixed-path baselines drop by about 1-2%. The paper also claims that the ring traversal is rotation-stable because rotations induce only cyclic shifts of ring loops.

Significance. If the empirical claims hold, the paper would make a useful contribution to Vision SSM design: it identifies scan order as a nontrivial factor, proposes a concrete traversal scheme with a plausible efficiency/accuracy trade-off, and introduces a simple channel-routing plug-in that appears to transfer to several backbones. The manuscript is clearly structured, defines its architecture in explicit equations, and compares against many recent Vision-Mamba variants. However, the central rotation-robustness mechanism is not established for the actual test-time transformation used in the experiments, and several experimental details (undefined evaluation split, missing error bars, inconsistent model specifications, absent supplementary material) prevent verification. The significance is therefore potential rather than demonstrated at this stage.

major comments (5)
  1. [Sec. 3.3, Eq. (2) and rotation behavior paragraph] The claim that in-plane rotations leave each pixel's ring index unchanged and that a ring's loop experiences only a cyclic shift is only true for a continuous rotation on an unbounded plane. In the experiments of Sec. 4.2, rotations of 30/60 degrees are rendered on a fixed canvas, which involves interpolation and padding. Pixels from outside the original disk enter rings, ring membership can change after discretization, and the ring content is not a cyclic shift of the original. No derivation or controlled experiment is provided to quantify this gap. The measured flat accuracy could instead result from training augmentation or from the averaging in Eq. (5) washing out local changes. Because rotation robustness is a headline contribution, the paper needs either a formal statement of the invariance under the actual transformation or an ablation that directly measures ring-index overlap and
  2. [Sec. 4.3, Table 3] The COCO results are reported on 'MS COCO (mini-val, 1×schedule, 1280×800)', but 'mini-val' is never defined. The standard COCO benchmark uses val2017; a non-standard split would make the numbers not directly comparable to the cited baselines. Please specify how the mini-val split is constructed, how many images it contains, and whether the same images are used for all models. If available, also report standard val2017 numbers for the main comparison.
  3. [Table 6 vs. Table 1; Sec. 4.6] There is an internal inconsistency: Table 1 reports PRISMamba at 22M parameters and 84.5% Top-1, while Table 6 lists 'PRISMamba 30M' at 84.5% and 'PRISMamba 50M' at 85.3% with no FLOPs/throughput. The Table 6 caption refers to 'Ring-Mamba' as a plug-in, but the rows are labeled PRISMamba. Moreover, Sec. 4.6 attributes robustness to 'mask-gated updates', but no mask-gating mechanism is defined in Section 3. These inconsistencies make the occlusion experiment impossible to interpret and need to be resolved.
  4. [Sec. 4.2, Table 2; Sec. 4.4, Table 4] All top-1 differences are reported as point estimates without error bars, seeds, or significance tests. Several claimed effects are small (e.g., +0.2-0.4 pp for PCF, and the 0.1-0.2 pp rotation drops for PRISMamba). Given typical run-to-run variance in ImageNet training, the reader cannot tell whether these differences are real. Please provide at least three independent runs with mean±std (or confidence intervals) for the main comparisons, especially the rotation table and the PCF abalation.
  5. [Sec. 1 and Sec. 3.2; Fig. 4] The paper claims 'the first systematic study of how traversal paths shape spatial adjacency in Vision SSMs', but the actual scan-order comparison (orders S1-S21 in Fig. 4) is deferred to a Supplementary Material that is not included. As submitted, there is no quantitative evidence for the central motivating claim that scan order matters beyond the ring-vs-fixed-path rotation table. Please include the scan-order ablation in the main text or an appendix, or soften the contribution claim.
minor comments (5)
  1. [Sec. 3.3, Eq. (2)] The formula for r(u,v) appears to be missing a norm symbol; as printed it reads as a 2-tuple. Also define the ring width Δr, the ring center, and the number of rings used in the experiments; without these values the architecture is not reproducible.
  2. [Sec. 4.2] Please specify how the rotated test set is constructed: which interpolation method, what padding mode, and whether rotation is applied at inference only. This is necessary for replicating the rotation robustness result.
  3. [Table 6] The mask column headers 'Mask 4×4' and 'Mask 16×16' are described in the caption as 'one of 2×2 tiles' and 'one of 4×4 tiles'. Are the numbers pixel sizes or tile counts? The notation is confusing and should be clarified.
  4. [Table 1 caption and various rows] There are typographical errors: 'measured with on Nvidia A100 GPU' should be 'measured on an Nvidia A100 GPU'. Some entries have spaces around 'G' inconsistently (e.g., '5.1 (-0.5) G'), and several references have non-standard capitalization (e.g., 'peng gao').
  5. [Fig. 1] The label 'Vmamba Zigma MaIR LocalMamba' at the top of the figure is not formatted as a list and does not clearly map to the different scan paths shown. Please revise the figure caption and labels for clarity.

Circularity Check

0 steps flagged

No significant circularity: architectural definitions and external benchmarks drive the claims; the rotation mechanism is under-supported but not a circular derivation.

full rationale

The paper's claimed derivation chain is: scan order affects spatial adjacency (Sec. 3.2); pixels are grouped by Euclidean distance into rings (Eq. 2); each ring is processed by a loop SSM (Eq. 4) and summarized by averaging outputs (Eq. 5); context propagates radially (Eq. 6); features are written back and fused (Eqs. 9-10). These equations define the architecture rather than fitting parameters to the reported outcomes. The headline numbers (ImageNet 84.5% Top-1, COCO 48.9 AP box / 43.2 AP mask, rotation stress table) are external empirical measurements against public benchmarks and published baselines; no parameter is fitted from the target metric and then presented as a prediction. The rotation-robustness assertion in Sec. 3.3 that 'In-plane rotations leave each pixel's ring index unchanged' is a mathematical consequence of Eq. (2) under exact rotations, not a circular reduction; it also does not by itself force the measured robustness because the experiments use rendered 30°/60° rotations with interpolation and padding, where the cyclic-shift premise is not exact. This is a correctness/support gap, not a self-referential fit. The Limitations section itself flags that 'severe rotations that create large padded regions still reduce valid information.' There are no load-bearing self-citations: the cited prior work is external, and no uniqueness theorem or ansatz is imported from the authors' own earlier results. Therefore no circular step can be exhibited, and the appropriate score is 0.

Axiom & Free-Parameter Ledger

4 free parameters · 3 axioms · 2 invented entities

The method borrows the standard Mamba recurrence and adds new ring and channel-routing structures. The paper does not prove these structures preserve the information needed for recognition; the central claim therefore rests on empirical validation plus several unstated modeling choices (ring width, ring origin, radial propagation order).

free parameters (4)
  • Ring width Δr = not reported
    Controls the number of rings R* and the length of ring loops and the radial chain; hand-chosen. The paper only states that practical choices make R* ≪ sqrt(HW), but never reports the value or a sensitivity study.
  • Ring center (cx, cy) = image center
    Fixed at the image center unless an object detector provides an alternative. The authors admit in Limitations that this may be suboptimal for off-center subjects or extreme aspect ratios.
  • Channel retention threshold = mean of |GAP| per layer
    The hard routing in Eq. (7) uses the layer-wise mean of absolute GAP values as the threshold. This is data-dependent rather than a fitted scalar, but the choice of mean over median is arbitrary and no comparison is reported.
  • Radial SSM capacity = not specified
    The dimensions and depth of the radial selective SSM in Eq. (6) are not reported; it is a free design choice that affects accuracy and cost.
axioms (3)
  • domain assumption The Mamba-style recurrence (Eq. 1) is a sufficient sequence operator for vision
    Imported from the SSM literature; the paper does not re-derive or question it.
  • ad hoc to paper Order-agnostic averaging of ring outputs (Eq. 5) preserves recognizability
    No theoretical justification is given; the method is validated only empirically on ImageNet/COCO.
  • ad hoc to paper Inner-to-outer radial propagation (Eq. 6) is a meaningful axis for global context
    The paper assumes that reading rings from center outward composes information correctly without a full 2D scan; this is not proven.
invented entities (2)
  • Concentric Ring Scan partition no independent evidence
    purpose: Rotation-robust serialization of 2D features for SSMs
    Introduced in §3.3; the only evidence is the paper's own experiments. No external prediction or formal guarantee distinguishes it from other scan orders.
  • Partial Channel Filtering (hard channel routing) no independent evidence
    purpose: Skip low-salience channels in the recurrent path to reduce compute and sharpen descriptors
    Introduced in §3.5; evidence is confined to in-paper ablations and plug-in tests on other backbones.

pith-pipeline@v1.3.0-alltime-deepseek · 17013 in / 14353 out tokens · 153791 ms · 2026-08-03T04:41:44.187724+00:00 · methodology

0 comments
read the original abstract

State Space Models (SSMs) have emerged as efficient alternatives to attention for vision tasks, offering lineartime sequence processing with competitive accuracy. Vision SSMs, however, require serializing 2D images into 1D token sequences along a predefined scan order, a factor often overlooked. We show that scan order critically affects performance by altering spatial adjacency, fracturing object continuity, and amplifying degradation under geometric transformations such as rotation. We present Partial RIng Scan Mamba (PRISMamba), a rotation-robust traversal that partitions an image into concentric rings, performs order-agnostic aggregation within each ring, and propagates context across rings through a set of short radial SSMs. Efficiency is further improved via partial channel filtering, which routes only the most informative channels through the recurrent ring pathway while keeping the rest on a lightweight residual branch. On ImageNet-1K, PRISMamba achieves 84.5% Top-1 with 3.9G FLOPs and 3,054 img/s on A100, outperforming VMamba in both accuracy and throughput while requiring fewer FLOPs. It also maintains performance under rotation, whereas fixed-path scans drop by 1~2%. These results highlight scan-order design, together with channel filtering, as a crucial, underexplored factor for accuracy, efficiency, and rotation robustness in Vision SSMs. Code will be released upon acceptance.

Figures

Figures reproduced from arXiv: 2602.04170 by Jun-Wei Hsieh, Kuan-Chuan Peng, Ming-Ching Chang, Xin Li, Yi-Kuan Hsieh, Yu-Chee Tseng.

Figure 1
Figure 1. Figure 1: Scanning order affects Vision-Mamba perfor￾mance. (a) Fixed-path scans (e.g., raster or serpentine in VMamba [37], Zigma [26], MaIR [32], LocalMamba [28]) pre￾serve sequence–space alignment only under flips. An in-plane ro￾tation (here 60◦ ) causes padding and global reindexing, fracturing the path so the recurrent kernel moves along misaligned neighbor￾hoods. (b) Our Ring Scan treats serialization as orde… view at source ↗
Figure 2
Figure 2. Figure 2: Architecture with Partial RIng Scan Mamba (PRISMamba). The image is patchified and processed by a four-stage backbone; stage i stacks Li PRISM blocks (Partial RIng Scan Mamba) with Ci output channels, and stages are separated by downsampling. Each PRISM performs order-agnostic aggregation over a subset of concentric rings (partial ring scan), composes information radially with a short sequence operator, an… view at source ↗
Figure 3
Figure 3. Figure 3: Ring Scan. Pixels are partitioned into concentric rings, which are interactively traversed in a clockwise or counterclockwise sequence. The resulting features are then aggregated in an order-independent fashion, proceeding from inner to outer rings. S1 S2 S3 S4 S5 S6 S7 S8 S9 S10 S11 S12 S 1 S 2 S 3 S 4 S 5 S 6 S 7 S 8 S 9 S 10 S 11 S 12 S 13 S 14 S 15 S 16 S 17 S 18 S 19 S 20 S 21 [PITH_FULL_IMAGE:figure… view at source ↗
Figure 4
Figure 4. Figure 4: Primitive scan orders. Twelve canonical paths (S1–S12) such as left-to-right raster, serpentine, and diagonal produce distinct 1D sequences from the same image. S1-12 evaluate single scans; S13–18 evaluate pairs of scans; S 19–21 aggregate four scans, enabling a systematic comparison of scan-order effects. In real implementations, if an object detector is performed first, (cx, cy) can be an object’s center… 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

84 extracted references · 23 linked inside Pith

  1. [1]

    Xcit: Cross-covariance image transformers.Advances in neural information processing systems, 34:20014–20027, 2021

    Alaaeldin Ali, Hugo Touvron, Mathilde Caron, Piotr Bo- janowski, Matthijs Douze, Armand Joulin, Ivan Laptev, Na- talia Neverova, Gabriel Synnaeve, Jakob Verbeek, et al. Xcit: Cross-covariance image transformers.Advances in neural information processing systems, 34:20014–20027, 2021. 3

    2021

  2. [2]

    Autoaugment: Learning augmentation strategies from data

    Ekin D Cubuk, Barret Zoph, Dandelion Mane, Vijay Vasude- van, and Quoc V Le. Autoaugment: Learning augmentation strategies from data. InProceedings of the IEEE/CVF con- ference on computer vision and pattern recognition, pages 113–123, 2019. 6

    2019

  3. [3]

    Deformable convolutional networks

    Jifeng Dai, Haozhi Qi, Yuwen Xiong, Yi Li, Guodong Zhang, Han Hu, and Yichen Wei. Deformable convolutional networks. InProceedings of the IEEE international confer- ence on computer vision, pages 764–773, 2017. 3

    2017

  4. [4]

    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

    2021

  5. [5]

    Flashattention-2: Faster attention with bet- ter parallelism and work partitioning.arXiv preprint arXiv:2307.08691, 2023

    Tri Dao. Flashattention-2: Faster attention with bet- ter parallelism and work partitioning.arXiv preprint arXiv:2307.08691, 2023. 3

    Pith/arXiv arXiv 2023

  6. [6]

    Flashattention: Fast and memory-efficient exact attention with io-awareness.Advances in neural information processing systems, 35:16344–16359, 2022

    Tri Dao, Dan Fu, Stefano Ermon, Atri Rudra, and Christo- pher R ´e. Flashattention: Fast and memory-efficient exact attention with io-awareness.Advances in neural information processing systems, 35:16344–16359, 2022. 3

    2022

  7. [7]

    Davit: Dual attention vision transform- ers

    Mingyu Ding, Bin Xiao, Noel Codella, Ping Luo, Jingdong Wang, and Lu Yuan. Davit: Dual attention vision transform- ers. InEuropean conference on computer vision, pages 74–

  8. [8]

    Scaling up your kernels to 31x31: Revisiting large kernel design in cnns

    Xiaohan Ding, Xiangyu Zhang, Jungong Han, and Guiguang Ding. Scaling up your kernels to 31x31: Revisiting large kernel design in cnns. InProceedings of the IEEE/CVF con- ference on computer vision and pattern recognition, pages 11963–11975, 2022. 3

    2022

  9. [9]

    Cswin transformer: A general vision transformer backbone with cross-shaped windows

    Xiaoyi Dong, Jianmin Bao, Dongdong Chen, Weiming Zhang, Nenghai Yu, Lu Yuan, Dong Chen, and Baining Guo. Cswin transformer: A general vision transformer backbone with cross-shaped windows. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12124–12134, 2022. 3

    2022

  10. [10]

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

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Syl- vain Gelly, et al. An image is worth 16x16 words: Trans- formers for image recognition at scale.arXiv preprint arXiv:2010.11929, 2020. 3

    Pith/arXiv arXiv 2010

  11. [11]

    RMT: Retentive Networks Meet Vision Trans- formers

    Qihang Fan, Huaibo Huang, Mingrui Chen, Hongmin Liu, and Ran He. RMT: Retentive Networks Meet Vision Trans- formers . In2024 IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition (CVPR), pages 5641–5651,

  12. [12]

    Hungry hungry hippos: To- wards language modeling with state space models.arXiv preprint arXiv:2212.14052, 2022

    Daniel Y Fu, Tri Dao, Khaled K Saab, Armin W Thomas, Atri Rudra, and Christopher R´e. Hungry hungry hippos: To- wards language modeling with state space models.arXiv preprint arXiv:2212.14052, 2022. 3

    Pith/arXiv arXiv 2022

  13. [13]

    Container: Context aggregation net- works

    peng gao, Jiasen Lu, hongsheng Li, Roozbeh Mottaghi, and Aniruddha Kembhavi. Container: Context aggregation net- works. InAdvances in Neural Information Processing Sys- tems, pages 19160–19171. Curran Associates, Inc., 2021. 3

    2021

  14. [14]

    Mamba: Linear-time sequence mod- eling with selective state spaces

    Albert Gu and Tri Dao. Mamba: Linear-time sequence mod- eling with selective state spaces. InFirst Conference on Lan- guage Modeling, 2024. 1, 3

    2024

  15. [15]

    Hippo: Recurrent memory with optimal polynomial projections.Advances in neural information processing sys- tems, 33:1474–1487, 2020

    Albert Gu, Tri Dao, Stefano Ermon, Atri Rudra, and Christo- pher R´e. Hippo: Recurrent memory with optimal polynomial projections.Advances in neural information processing sys- tems, 33:1474–1487, 2020. 3

    2020

  16. [16]

    Efficiently modeling long sequences with structured state spaces.arXiv preprint arXiv:2111.00396, 2021

    Albert Gu, Karan Goel, and Christopher R ´e. Efficiently modeling long sequences with structured state spaces.arXiv preprint arXiv:2111.00396, 2021. 1, 3

    Pith/arXiv arXiv 2021

  17. [17]

    On the parameterization and initialization of diagonal state space models.Advances in Neural Information Processing Systems, 35:35971–35983, 2022

    Albert Gu, Karan Goel, Ankit Gupta, and Christopher R ´e. On the parameterization and initialization of diagonal state space models.Advances in Neural Information Processing Systems, 35:35971–35983, 2022. 3

    2022

  18. [18]

    Diagonal state spaces are as effective as structured state spaces.Advances in neural information processing systems, 35:22982–22994,

    Ankit Gupta, Albert Gu, and Jonathan Berant. Diagonal state spaces are as effective as structured state spaces.Advances in neural information processing systems, 35:22982–22994,

  19. [19]

    On the connection between local attention and dynamic depth-wise convolution.arXiv preprint arXiv:2106.04263, 2021

    Qi Han, Zejia Fan, Qi Dai, Lei Sun, Ming-Ming Cheng, Ji- aying Liu, and Jingdong Wang. On the connection between local attention and dynamic depth-wise convolution.arXiv preprint arXiv:2106.04263, 2021. 3

    Pith/arXiv arXiv 2021

  20. [20]

    Flatten wisely: How patch order shapes mamba-powered vision for mri segmentation.arXiv preprint arXiv:2507.13384, 2025

    Osama Hardan, Omar Elshenhabi, Tamer Khattab, and Mo- hamed Mabrok. Flatten wisely: How patch order shapes mamba-powered vision for mri segmentation.arXiv preprint arXiv:2507.13384, 2025. 2 9

    Pith/arXiv arXiv 2025

  21. [21]

    Liquid structural state-space models.arXiv preprint arXiv:2209.12951, 2022

    Ramin Hasani, Mathias Lechner, Tsun-Hsuan Wang, Makram Chahine, Alexander Amini, and Daniela Rus. Liquid structural state-space models.arXiv preprint arXiv:2209.12951, 2022. 3

    Pith/arXiv arXiv 2022

  22. [22]

    Global context vision transformers

    Ali Hatamizadeh, Hongxu Yin, Greg Heinrich, Jan Kautz, and Pavlo Molchanov. Global context vision transformers. In International conference on machine learning, pages 12633– 12646. PMLR, 2023. 3

    2023

  23. [23]

    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. 3

    2016

  24. [24]

    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. 3

    Pith/arXiv arXiv 2017

  25. [25]

    Squeeze-and-excitation net- works

    Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation net- works. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 7132–7141, 2018. 5, 6, 8

    2018

  26. [26]

    Zigma: A dit-style zigzag mamba diffusion model

    Vincent Tao Hu, Stefan Andreas Baumann, Ming Gui, Olga Grebenkova, Pingchuan Ma, Johannes Fischer, and Bj ¨orn Ommer. Zigma: A dit-style zigzag mamba diffusion model. InEuropean conference on computer vision, pages 148–166. Springer, 2024. 1, 4, 6, 7

    2024

  27. [27]

    Densely connected convolutional net- works

    Gao Huang, Zhuang Liu, Laurens Van Der Maaten, and Kil- ian Q Weinberger. Densely connected convolutional net- works. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 4700–4708, 2017. 3

    2017

  28. [28]

    Localmamba: Visual state space model with windowed selective scan

    Tao Huang, Xiaohuan Pei, Shan You, Fei Wang, Chen Qian, and Chang Xu. Localmamba: Visual state space model with windowed selective scan. InEuropean Conference on Com- puter Vision, pages 12–22. Springer, 2024. 1, 6, 7

    2024

  29. [29]

    yolov11.https://github.com/ultralytics, 2024

    Glenn Jocher. yolov11.https://github.com/ultralytics, 2024. 2

    2024

  30. [30]

    Transformers are rnns: Fast autoregressive transformers with linear attention

    Angelos Katharopoulos, Apoorv Vyas, Nikolaos Pappas, and Franc ¸ois Fleuret. Transformers are rnns: Fast autoregressive transformers with linear attention. InInternational confer- ence on machine learning, pages 5156–5165. PMLR, 2020. 3

    2020

  31. [31]

    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. 3

    2012

  32. [32]

    Mair: A locality-and continuity- preserving mamba for image restoration

    Boyun Li, Haiyu Zhao, Wenxin Wang, Peng Hu, Yuan- biao Gou, and Xi Peng. Mair: A locality-and continuity- preserving mamba for image restoration. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 7491–7501, 2025. 1, 4, 6, 7

    2025

  33. [33]

    Scale-aware modulation meet transformer

    Weifeng Lin, Ziheng Wu, Jiayu Chen, Jun Huang, and Lian- wen Jin. Scale-aware modulation meet transformer. In 2023 IEEE/CVF International Conference on Computer Vi- sion (ICCV), pages 5992–6003, 2023. 3

    2023

  34. [34]

    Defmamba: Deformable vi- sual state space model

    Leiye Liu, Miao Zhang, Jihao Yin, Tingwei Liu, Wei Ji, Yongri Piao, and Huchuan Lu. Defmamba: Deformable vi- sual state space model. InProceedings of the Computer Vi- sion and Pattern Recognition Conference, pages 8838–8847,

  35. [35]

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

    Shiwei Liu, Tianlong Chen, Xiaohan Chen, Xuxi Chen, Qiao Xiao, Boqian Wu, Tommi K¨arkk¨ainen, Mykola Pechenizkiy, Decebal Mocanu, and Zhangyang Wang. More convnets in the 2020s: Scaling up kernels beyond 51x51 using sparsity. arXiv preprint arXiv:2207.03620, 2022. 3

    Pith/arXiv arXiv 2022

  36. [36]

    Vision mamba: A compre- hensive survey and taxonomy.IEEE Transactions on Neural Networks and Learning Systems, 2025

    Xiao Liu, Chenxu Zhang, Fuxiang Huang, Shuyin Xia, Guoyin Wang, and Lei Zhang. Vision mamba: A compre- hensive survey and taxonomy.IEEE Transactions on Neural Networks and Learning Systems, 2025. 4

    2025

  37. [37]

    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, 4, 6, 7

    2024

  38. [38]

    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. 3

    2021

  39. [39]

    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,

  40. [40]

    Sgdr: Stochas- tic gradient descent with warm restarts.arXiv preprint arXiv:1608.03983, 2016

    Ilya Loshchilov and Frank Hutter. Sgdr: Stochas- tic gradient descent with warm restarts.arXiv preprint arXiv:1608.03983, 2016. 6

    Pith/arXiv arXiv 2016

  41. [41]

    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

  42. [42]

    Sparx: A sparse cross-layer connection mechanism for hierarchical vision mamba and transformer networks

    Meng Lou, Yunxiang Fu, and Yizhou Yu. Sparx: A sparse cross-layer connection mechanism for hierarchical vision mamba and transformer networks. InProceedings of the AAAI Conference on Artificial Intelligence, pages 19104– 19114, 2025. 6, 7

    2025

  43. [43]

    Mega: moving average equipped gated atten- tion.arXiv preprint arXiv:2209.10655, 2022

    Xuezhe Ma, Chunting Zhou, Xiang Kong, Junxian He, Liangke Gui, Graham Neubig, Jonathan May, and Luke Zettlemoyer. Mega: moving average equipped gated atten- tion.arXiv preprint arXiv:2209.10655, 2022. 3

    Pith/arXiv arXiv 2022

  44. [44]

    Tinyvim: Frequency decoupling for tiny hybrid vision mamba

    Xiaowen Ma, Zhenliang Ni, and Xinghao Chen. Tinyvim: Frequency decoupling for tiny hybrid vision mamba. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 23519–23529, 2025. 4

    2025

  45. [45]

    Long range language modeling via gated state spaces.arXiv preprint arXiv:2206.13947, 2022

    Harsh Mehta, Ankit Gupta, Ashok Cutkosky, and Behnam Neyshabur. Long range language modeling via gated state spaces.arXiv preprint arXiv:2206.13947, 2022. 3

    Pith/arXiv arXiv 2022

  46. [46]

    Simba: Simplified mamba-based architecture for vision and multivariate time series.arXiv preprint arXiv:2403.15360, 2024

    Badri N Patro and Vijay S Agneeswaran. Simba: Simplified mamba-based architecture for vision and multivariate time series.arXiv preprint arXiv:2403.15360, 2024. 6, 7

    Pith/arXiv arXiv 2024

  47. [47]

    Efficientvmamba: Atrous selective scan for light weight visual mamba

    Xiaohuan Pei, Tao Huang, and Chang Xu. Efficientvmamba: Atrous selective scan for light weight visual mamba. InPro- ceedings of the AAAI Conference on Artificial Intelligence, pages 6443–6451, 2025. 4, 6, 7

    2025

  48. [48]

    Rwkv: 10 Reinventing rnns for the transformer era.arXiv preprint arXiv:2305.13048, 2023

    Bo Peng, Eric Alcaide, Quentin Anthony, Alon Albalak, Samuel Arcadinho, Stella Biderman, Huanqi Cao, Xin Cheng, Michael Chung, Matteo Grella, et al. Rwkv: 10 Reinventing rnns for the transformer era.arXiv preprint arXiv:2305.13048, 2023. 3

    Pith/arXiv arXiv 2023

  49. [49]

    Designing network design spaces

    Ilija Radosavovic, Raj Prateek Kosaraju, Ross Girshick, Kaiming He, and Piotr Doll ´ar. Designing network design spaces. InProceedings of the IEEE/CVF conference on com- puter vision and pattern recognition, pages 10428–10436,

  50. [50]

    Hornet: Efficient high- order spatial interactions with recursive gated convolutions

    Yongming Rao, Wenliang Zhao, Yansong Tang, Jie Zhou, Ser Nam Lim, and Jiwen Lu. Hornet: Efficient high- order spatial interactions with recursive gated convolutions. Advances in Neural Information Processing Systems, 35: 10353–10366, 2022. 3

    2022

  51. [51]

    Vm- unet: Vision mamba unet for medical image segmentation

    Jiacheng Ruan, Jincheng Li, and Suncheng Xiang. Vm- unet: Vision mamba unet for medical image segmentation. ACM Transactions on Multimedia Computing, Communica- tions and Applications, 2024. 4

    2024

  52. [52]

    Groupmamba: Effi- cient group-based visual state space model

    Abdelrahman Shaker, Syed Talal Wasim, Salman Khan, Juergen Gall, and Fahad Shahbaz Khan. Groupmamba: Effi- cient group-based visual state space model. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 14912–14922, 2025. 6, 7

    2025

  53. [53]

    Transnext: Robust foveal visual perception for vi- sion transformers

    Dai Shi. Transnext: Robust foveal visual perception for vi- sion transformers. In2024 IEEE/CVF Conference on Com- puter Vision and Pattern Recognition (CVPR), pages 17773– 17783, 2024. 3

    2024

  54. [54]

    Vssd: Vision mamba with non-causal state space duality

    Yuheng Shi, Mingjia Li, Minjing Dong, and Chang Xu. Vssd: Vision mamba with non-causal state space duality. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10819–10829, 2025. 4, 6, 7

    2025

  55. [55]

    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. 3

    Pith/arXiv arXiv 2014

  56. [56]

    Simplified state space layers for sequence modeling

    Jimmy TH Smith, Andrew Warrington, and Scott W Linder- man. Simplified state space layers for sequence modeling. arXiv preprint arXiv:2208.04933, 2022. 3

    Pith/arXiv arXiv 2022

  57. [57]

    Retentive network: A successor to transformer for large language mod- els.arXiv preprint arXiv:2307.08621, 2023

    Yutao Sun, Li Dong, Shaohan Huang, Shuming Ma, Yuqing Xia, Jilong Xue, Jianyong Wang, and Furu Wei. Retentive network: A successor to transformer for large language mod- els.arXiv preprint arXiv:2307.08621, 2023. 3

    Pith/arXiv arXiv 2023

  58. [58]

    Going deeper with convolutions

    Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 1–9, 2015. 3

    2015

  59. [59]

    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,

  60. [60]

    Integrally pre-trained transformer pyramid net- works

    Yunjie Tian, Lingxi Xie, Zhaozhi Wang, Longhui Wei, Xi- aopeng Zhang, Jianbin Jiao, Yaowei Wang, Qi Tian, and Qixiang Ye. Integrally pre-trained transformer pyramid net- works. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18610– 18620, 2023. 3

    2023

  61. [61]

    Yolov12: Attention-centric real-time object detectors.arXiv preprint arXiv:2502.12524, 2025

    Yunjie Tian, Qixiang Ye, and David Doermann. Yolov12: Attention-centric real-time object detectors.arXiv preprint arXiv:2502.12524, 2025. 2

    Pith/arXiv arXiv 2025

  62. [62]

    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. 3, 6

    2021

  63. [63]

    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. 3

    2017

  64. [64]

    Mamba-reg: Vision mamba also needs registers

    Feng Wang, Jiahao Wang, Sucheng Ren, Guoyizhe Wei, Jieru Mei, Wei Shao, Yuyin Zhou, Alan Yuille, and Cihang Xie. Mamba-reg: Vision mamba also needs registers. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 14944–14953, 2025. 4

    2025

  65. [65]

    Adventurer: Optimizing vision mamba architecture designs for efficiency

    Feng Wang, Timing Yang, Yaodong Yu, Sucheng Ren, Guoyizhe Wei, Angtian Wang, Wei Shao, Yuyin Zhou, Alan Yuille, and Cihang Xie. Adventurer: Optimizing vision mamba architecture designs for efficiency. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 30157–30166, 2025. 4, 6, 7

    2025

  66. [66]

    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. 3

    2021

  67. [67]

    Cbam: Convolutional block attention module

    Sanghyun Woo, Jongchan Park, Joon-Young Lee, and In So Kweon. Cbam: Convolutional block attention module. In Proceedings of the European conference on computer vision (ECCV), pages 3–19, 2018. 5, 6, 8

    2018

  68. [68]

    Boosting vision state space model with fractal scanning

    Haoke Xiao, Lv Tang, Peng-tao Jiang, Hao Zhang, Jinwei Chen, and Bo Li. Boosting vision state space model with fractal scanning. InProceedings of the AAAI Conference on Artificial Intelligence, pages 8646–8654, 2025. 4, 6, 7

    2025

  69. [69]

    Quadmamba: Learning quadtree-based selective scan for vi- sual state space model.Advances in Neural Information Pro- cessing Systems, 37:117682–117707, 2024

    Fei Xie, Weijia Zhang, Zhongdao Wang, and Chao Ma. Quadmamba: Learning quadtree-based selective scan for vi- sual state space model.Advances in Neural Information Pro- cessing Systems, 37:117682–117707, 2024. 6, 7

    2024

  70. [70]

    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. 2, 4, 6, 7

    Pith/arXiv arXiv 2024

  71. [71]

    Focal self-attention for local-global interactions in vision transformers.arXiv preprint arXiv:2107.00641, 2021

    Jianwei Yang, Chunyuan Li, Pengchuan Zhang, Xiyang Dai, Bin Xiao, Lu Yuan, and Jianfeng Gao. Focal self-attention for local-global interactions in vision transformers.arXiv preprint arXiv:2107.00641, 2021. 3

    Pith/arXiv arXiv 2021

  72. [72]

    Gated linear attention trans- formers with hardware-efficient training.arXiv preprint arXiv:2312.06635, 2023

    Songlin Yang, Bailin Wang, Yikang Shen, Rameswar Panda, and Yoon Kim. Gated linear attention trans- formers with hardware-efficient training.arXiv preprint arXiv:2312.06635, 2023. 3

    Pith/arXiv arXiv 2023

  73. [73]

    Selective visual prompting in vision mamba

    Yifeng Yao, Zichen Liu, Zhenyu Cui, Yuxin Peng, and Ji- ahuan Zhou. Selective visual prompting in vision mamba. InProceedings of the AAAI Conference on Artificial Intelli- gence, pages 22083–22091, 2025. 4

    2025

  74. [74]

    Shvit: Single-head vision transformer with memory efficient macro design

    Seokju Yun and Youngmin Ro. Shvit: Single-head vision transformer with memory efficient macro design. InPro- 11 ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5756–5767, 2024. 3

    2024

  75. [75]

    mixup: Beyond empirical risk minimiza- tion.arXiv preprint arXiv:1710.09412, 2017

    Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimiza- tion.arXiv preprint arXiv:1710.09412, 2017. 6

    Pith/arXiv arXiv 2017

  76. [76]

    A survey on visual mamba.Applied Sciences, 14(13):5683, 2024

    Hanwei Zhang, Ying Zhu, Dan Wang, Lijun Zhang, Tianx- iang Chen, Ziyang Wang, and Zi Ye. A survey on visual mamba.Applied Sciences, 14(13):5683, 2024. 4

    2024

  77. [77]

    2dmamba: Efficient state space model for image repre- sentation with applications on giga-pixel whole slide image classification

    Jingwei Zhang, Anh Tien Nguyen, Xi Han, Vincent Quoc- Huy Trinh, Hong Qin, Dimitris Samaras, and Mahdi S Hos- seini. 2dmamba: Efficient state space model for image repre- sentation with applications on giga-pixel whole slide image classification. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 3583–3592, 2025. 4

    2025

  78. [78]

    Mamba as a bridge: Where vision foundation models meet vision language models for domain-generalized semantic segmentation

    Xin Zhang and Robby T Tan. Mamba as a bridge: Where vision foundation models meet vision language models for domain-generalized semantic segmentation. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 14527–14537, 2025. 4

    2025

  79. [79]

    Hivit: A simpler and more efficient design of hierarchical vision transformer

    Xiaosong Zhang, Yunjie Tian, Lingxi Xie, Wei Huang, Qi Dai, Qixiang Ye, and Qi Tian. Hivit: A simpler and more efficient design of hierarchical vision transformer. InThe eleventh international conference on learning representa- tions, 2023. 3

    2023

  80. [80]

    Graformer: Graph-oriented transformer for 3d pose estimation

    Weixi Zhao, Weiqiang Wang, and Yunjie Tian. Graformer: Graph-oriented transformer for 3d pose estimation. InPro- ceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 20438–20447, 2022. 3

    2022

Showing first 80 references.