Pith. sign in

REVIEW 4 major objections 5 minor 14 references

Mamba-MOC: A Multicategory Remote Object Counting via State Space Model

T0 review · 4 major / 5 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read Mamba-MOC shows that a Mamba backbone with cross-scale and local-context modules sets new accuracy records on a multicategory remote-object-counting benchmark.

desk verdict Plausible Mamba-based counting network, but the SOTA claim rests on a Table I with arithmetic errors and unverified baseline settings. read the letter →

arxiv 2501.06697 v2 pith:BND7LSV2 submitted 2025-01-12 cs.CV cs.AI

Peng Liu , Sen Lei , Heng-Chao Li This is my paper · ORCID
classification cs.CVcs.AI
keywords remotesensingobjectcountingMambastate-spacemodelmulticategorydensitymapestimationcross-scaleinteractionNWPU-MOC
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 argues that selective state-space models, which process long sequences in linear time, can replace CNN and transformer encoders for multi-category object counting in remote sensing images. To show this, the authors build Mamba-MOC, a Vmamba-based network that adds two components: a cross-scale interaction module that fuses coarse and fine features, and a context state space model that injects local neighborhood information into the global scan. On the NWPU-MOC benchmark, Mamba-MOC reports the best overall MSE of 9.5794 and WMSE of 27.2012 among the compared methods, and the best per-category error in five of six categories. A sympathetic reader would take this as evidence that Mamba-style architectures are worth pursuing for remote counting tasks that need global context without quadratic attention cost.

What carries the argument

The load-bearing pieces are the Vmamba backbone, the Cross-Scale Interaction Module (CIM), and the Context State Space Model (CSSM). CIM resizes and concatenates multi-level features, runs them through a Mamba block, then splits and re-integrates them with a gating mechanism. CSSM builds on the observation that the output matrix $C$ in the state-space equation acts like a query: it forms $Q_l$ from multi-scale dilated convolutions, adds it to $C$, and adds a local feature $F_l$ to the output, so the scan keeps its linear-complexity global view while also seeing local neighborhood structure. This combination is what carries the reported accuracy gains.

What would settle it

Re-run the five compared baselines (CSRNet, SFCN, PSGCNet, DSACA, MCC) on NWPU-MOC under the paper's exact density-map protocol, a Gaussian kernel of bandwidth 4 and size 15, and the official 2,391/1,025 train/test split; if any baseline reaches an MSE at or below 9.5794, or if the reproduced numbers differ materially from Table I, the state-of-the-art claim would be overturned.

Watch

Extended reading notes

Core claim

The central claim is that Mamba, originally designed for linear-time sequence modeling, can be adapted into a strong encoder for multicategory remote object counting by solving two problems: integrating features across scales and compensating for the causal-scan limitation of selective state-space models on 2D images. The paper's network uses a Vmamba backbone to extract hierarchical features, a cross-scale interaction module to align and fuse them, and a context state space model whose output equation becomes $y(t) = (C + Q_l)h(t) + F_l$, where $Q_l$ is a multi-scale local query and $F_l$ is a local feature. With these additions, the method improves on the compared CNN- and transformer-based counters on NWPU-MOC, reducing the inter-category average MSE and weighted MSE to 9.5794 and 27.2012, respectively.

Load-bearing premise

The comparison is only valid if every baseline was trained and tested with the same Gaussian-kernel density maps (bandwidth 4, size 15) and the same 2,391/1,025 split, which the paper states for its own runs but does not confirm for the quoted baseline results.

Editorial extensions

If this is right

  • Mamba-MOC reports the best overall MSE (9.5794) and WMSE (27.2012) on NWPU-MOC among the compared methods, and the best per-category error in five of six categories.
  • Because the state-space backbone runs in linear time, the method offers a computationally cheaper route to global-context counting than transformer-based alternatives on large remote images.
  • The ablation study shows that adding the cross-scale interaction module lowers both errors, and adding the context state space model lowers them further, so each component contributes to the reported result.
  • The authors present this as the first application of Mamba to remote sensing object counting, which positions selective state-space models as a viable family for this task.

Reading between the lines

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

  • An extension the paper does not test: the local-context query trick in CSSM could transfer to other dense prediction tasks with causal-scan artifacts, such as segmentation or change detection, and might improve their fine-grained outputs.
  • If the headline margin survives re-running the baselines under identical density-map generation and split settings, a Mamba-based encoder is a practical substitute for transformers in very large remote scenes, where quadratic attention becomes the bottleneck.
  • The method is demonstrated on a single benchmark, so the general claim depends on replication; a natural next check is whether the same components transfer to other multicategory counting datasets or to counting across video frames.
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 / 5 minor

Summary. The manuscript proposes Mamba-MOC, a Mamba-based network for multicategory remote sensing object counting. The architecture combines a Vmamba backbone, a cross-scale interaction module (CIM) for fusing multi-level features, and two Context State Space (CSS) blocks built around a Context State Space Model (CSSM) that augments the SSM output matrix with local convolutional features. On the NWPU-MOC dataset the method reports a mean squared error (MSE) of 9.5794 and a weighted MSE (WMSE) of 27.2012, outperforming five baselines; an ablation study attributes gains to CIM and CSSM. The paper claims to be the first application of Mamba to remote sensing object counting and states that code will be released.

Significance. If the reported gains are reproducible under a common experimental protocol, the paper provides a useful demonstration that a Mamba-based architecture can reach competitive accuracy for multi-category remote sensing counting while retaining linear complexity; the CSSM module could also be reused by other dense prediction tasks. The strengths are the clear overall architecture, the use of an ablation study, and the commitment to release code. The significance is currently limited by the absence of a documented baseline protocol, by arithmetic inconsistencies in the main comparison table, and by the lack of statistical variability information; these issues leave the headline state-of-the-art claim conditional on a controlled reproduction.

major comments (4)
  1. [Section IV-A and Table I] The paper does not state whether the five baseline entries in Table I were produced under the same protocol as the proposed method: same ground-truth density-map generation (Gaussian kernel bandwidth 4, size 15), same input resolution 512 by 512, same optimizer and split, and same evaluation code. Because PSGCNet [10] and DSACA [11] were published before the NWPU-MOC dataset [12], their numbers cannot come from the original papers and must originate from an unstated reimplementation or third-party benchmark. The central state-of-the-art claim is therefore conditional on a controlled, code-level reproduction or on explicit documentation of the baseline protocol.
  2. [Table I] The overall MSE column contains arithmetic inconsistencies. For SFCN, the six per-category MSEs (10.5572, 13.8747, 8.3646, 7.3178, 32.2336, 0.4768) average to 12.1374, not the reported 14.8041. For MCC, the corresponding values (4.3597, 12.7091, 7.3538, 6.8366, 35.3594, 0.4768) average to 11.1826, not the reported 11.4826. The other four rows are exact category averages, so the discrepancy is not explained by the definition of the overall metric. The table should be corrected and the provenance of these baseline numbers verified.
  3. [Section III-C, Eq. (8)] Equation (8) is dimensionally unclear as written. In Eq. (1), C is an N by N matrix and h(t) is the hidden state, while in Eq. (7) Q_l has shape HW by D and F_l has shape H by W by C. The residual additions (C+Q_l)h(t) and y(t)=...+F_l therefore require reshaping, projection, or broadcasting steps that the text does not specify. Because this is the core mechanism of the proposed CSSM, the formulation is not reproducible from the paper in its current form. Please provide exact shapes and an ablation that separates the contribution of the local query addition from the local output addition.
  4. [Section IV-C, Table II] The ablation shows that adding CIM changes the baseline MSE from 10.3800 to 10.3533, a difference of 0.0267, and that adding CSSM then produces a much larger change. No repeated runs, seeds, or error bars are reported, so the small CIM gain may be within run-to-run noise. Given that the claimed advantage over some baselines is also small, the paper should report means and standard deviations over multiple runs, or otherwise justify that the differences are not noise.
minor comments (5)
  1. [Section IV-A] The formulas for MSE and WMSE are not given; since the paper relies on these metrics for its headline results, please define them explicitly rather than only citing [12].
  2. [Section IV-B and Table I] There are several typos: 'Tabel I' should be 'Table I', 'Countainer' should be 'Container', and the column header ordering 'MSE WMSE MAE MSE ...' is confusing and should be aligned with the data columns.
  3. [Section II, Eq. (1)] The dimensions in Eq. (1) are inconsistent: if h(t) is the hidden state, C should not be N by N in the standard SSM formulation; please clarify the intended shapes (e.g., C in R^{1 by N} or the equivalent for multi-channel outputs).
  4. [Section III-C] The text says 'As shown in Fig. 3' when describing the CSSM architecture, but Fig. 3 is the visualization results; the architecture is in Fig. 2. Please correct the cross-reference.
  5. [Section III-C] The phrase 'multi-scale contextual features' describes only two dilation rates (1 and 2); this is a minor terminology issue, but consider clarifying that the multi-scale aspect comes from the two receptive fields.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular dependency found: the method is an architectural contribution evaluated on an external benchmark, and the authors' self-citations are background only.

full rationale

The paper's central claim is an empirical state-of-the-art result on the external NWPU-MOC benchmark, obtained by training a Mamba-based counting network and comparing its reported MSE/WMSE with numbers from prior published methods. No equation in the paper defines a fitted quantity as a prediction, and no result is derived from its own output. The two self-citations ([4] and [6]) are used only as literature background for why global modeling is useful and where Mamba has been applied; they do not supply a load-bearing theorem, a uniqueness argument, or a mandated ansatz. The Context State Space Model is presented as a design directly from Equations (7)-(8), inspired by the cited MambaIRv2 attention interpretation, but this is an architectural motivation rather than a circular derivation. The ablation study in Table II isolates the proposed modules and reports monotone improvements, which is standard empirical evidence rather than self-verification. The only substantive concern is that the baseline numbers in Table I may come from different density-map or split protocols, since some baselines predate the dataset; however, that is a reproducibility and external-validity risk, not circularity. No circular step can be exhibited from the paper's own text, so the circularity score is 0.

Assumptions & free parameters 8 free parameters · 6 assumptions · 0 invented entities

The central empirical claim rests on the selected hyperparameters, the dataset protocol assumptions, and the unverified design assumption about the output matrix C acting as a query. None of these constitute a circular derivation, but several are load-bearing for the reported SOTA numbers.

free parameters (8)
  • Gaussian kernel bandwidth = 4
    Used for ground-truth density map generation in Section IV-A; all metrics depend on this choice and no sensitivity analysis is provided.
  • Gaussian kernel size = 15
    Together with bandwidth, fixes the spatial extent of each object annotation in the density map (Section IV-A).
  • Input training resolution = 512 x 512
    Set in Section IV-A; affects the receptive field and density map resolution.
  • Learning rate = 5e-5
    AdamW learning rate in Section IV-A; no schedule or sensitivity analysis is reported.
  • Weight decay = 1e-4
    AdamW weight decay in Section IV-A; chosen by hand.
  • Batch size = 8
    Section IV-A; standard choice, no sensitivity reported.
  • Training epochs = 200
    Total epochs in Section IV-A; no early stopping or variance across seeds is reported.
  • CSSM dilation rates = 1 and 2
    Dilation rates of the two 3x3 convolutions in Eq. 7; they define the multi-scale local context and are chosen manually.
assumptions (6)
  • standard math Zero-order hold discretization formulas for the SSM (Eq. 2) are correct and applicable.
    Borrowed from Mamba [7]; used to define the model, but not proven in this paper.
  • domain assumption The Vmamba backbone provides suitable multi-scale features for counting.
    Section III-A uses Vmamba as encoder; its suitability is inherited from prior classification work [5] and not justified for counting.
  • ad hoc to paper The output matrix C in the state space model acts as a query in attention, so adding a local-context feature Q_l to C improves counting.
    Section III-C invokes MambaIRv2 [14] for this claim; the paper provides no analysis or derivation for why the addition is valid.
  • domain assumption Fixed Gaussian density maps accurately represent object counts for training and evaluation.
    Section IV-A generates ground truth with kernel bandwidth 4 and size 15; assumes this matches the dataset protocol and that density-based counting is valid.
  • domain assumption Grouping the 14 NWPU-MOC categories into 6 coarse categories is a valid evaluation setup.
    Section IV-A follows [12]; if the grouping differs from benchmark convention, comparisons may not align.
  • domain assumption Prior methods' results in Table I were produced under the same training/test split and density-map generation.
    Section IV-B compares to numbers from [8] through [12]; the paper does not confirm protocol equivalence.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Mamba-MOC: A Multicategory Remote Object Counting via State Space Model." pith.science (2026). https://pith.science/paper/BND7LSV2

@misc{pith2026250106697,
  author       = {Pith},
  title        = {Pith review of: Mamba-MOC: A Multicategory Remote Object Counting via State Space Model},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BND7LSV2}},
  note         = {Machine review of arXiv:2501.06697}
}
read the original abstract

Multicategory remote object counting is a fundamental task in computer vision, aimed at accurately estimating the number of objects of various categories in remote images. Existing methods rely on CNNs and Transformers, but CNNs struggle to capture global dependencies, and Transformers are computationally expensive, which limits their effectiveness in remote applications. Recently, Mamba has emerged as a promising solution in the field of computer vision, offering a linear complexity for modeling global dependencies. To this end, we propose Mamba-MOC, a mamba-based network designed for multi-category remote object counting, which represents the first application of Mamba to remote sensing object counting. Specifically, we propose a cross-scale interaction module to facilitate the deep integration of hierarchical features. Then we design a context state space model to capture both global and local contextual information and provide local neighborhood information during the scan process. Experimental results in large-scale realistic scenarios demonstrate that our proposed method achieves state-of-the-art performance compared with some mainstream counting algorithms.

Figures

Figures reproduced from arXiv: 2501.06697 by the authors.

Figure 1
Figure 1. Overview of the proposed Mamba-MOC. parameters A and B from the continuous system into the discrete parameters A and B: A = exp(△A), B = (△A) −1 (exp(△A) − I) · △B (2) After that, the discretized model can be represented as: ht = Aht−1 + Bxt, yt = Cht (3) Finally, it can be represented as a global convolution, formally defined as: K = (CB, CAB, ..., CA L−1 B), y = x ∗ K (4) where ∗ denotes convolution operation, L i… view at source ↗
Figure 2
Figure 2. Architecture of the context state space model. [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Visualization results of Mamba-MOC on NWPU-MOC dataset. [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

14 extracted references · 10 canonical work pages

  1. [10]

    PSGCNet: A pyramidal scale and global context guided network for dense object counting in remote-sensing images,

    G. Gao, Q. Liu, Z. Hu, L. Li, Q. Wen, and Y . Wang, “PSGCNet: A pyramidal scale and global context guided network for dense object counting in remote-sensing images,” IEEE Transactions on Geoscience and Remote Sensing , vol. 60, pp. 1–12, 2022

    work page 2022

  2. [11]

    Dilated- scale-aware category-attention convnet for multi-class object counting,

    W. Xu, D. Liang, Y . Zheng, J. Xie, and Z. Ma, “Dilated- scale-aware category-attention convnet for multi-class object counting,” IEEE Signal Processing Letters , vol. 28, pp. 1570– 1574, 2021

    work page 2021

  3. [12]

    NWPU-MOC: A benchmark for fine-grained multicategory object counting in aerial images,

    J. Gao, L. Zhao, and X. Li, “NWPU-MOC: A benchmark for fine-grained multicategory object counting in aerial images,” IEEE Transactions on Geoscience and Remote Sensing , vol. 62, pp. 1–14, 2024

    work page 2024

  4. [1]

    Urban planning and building smart cities based on the internet of things using big data analytics,

    M. M. Rathore, A. Ahmad, A. Paul, and S. Rho, “Urban planning and building smart cities based on the internet of things using big data analytics,” Computer networks, vol. 101, pp. 63– 80, 2016

    work page 2016

  5. [2]

    Transformer for tree counting in aerial images,

    G. Chen and Y . Shang, “Transformer for tree counting in aerial images,” Remote Sensing , vol. 14, no. 3, p. 476, 2022

    work page 2022

  6. [3]

    Detecting and counting sheep with a convolutional neural network,

    F. Sarwar, A. Griffin, P. Periasamy, K. Portas, and J. Law, “Detecting and counting sheep with a convolutional neural network,” in 2018 15th IEEE International Conference on Advanced Video and Signal Based Surveillance (A VSS) . IEEE, 2018, pp. 1–6

    work page 2018

  7. [4]

    Ex- ploring fine-grained image-text alignment for referring remote sensing image segmentation,

    S. Lei, X. Xiao, T. Zhang, H.-C. Li, Z. Shi, and Q. Zhu, “Ex- ploring fine-grained image-text alignment for referring remote sensing image segmentation,” IEEE Transactions on Geoscience and Remote Sensing , pp. 1–1, 2024

    work page 2024

  8. [5]

    Vmamba: Visual state space model,

    Y . Liu, Y . Tian, Y . Zhao, H. Yu, L. Xie, Y . Wang, Q. Ye, and Y . Liu, “Vmamba: Visual state space model,” 2024. [Online]. Available: https://arxiv.org/abs/2401.10166

    arXiv 2024

Show all 14 references
  1. [6]

    FER-YOLO-Mamba: Facial expression detection and classification based on selective state space,

    H. Ma, S. Lei, T. Celik, and H.-C. Li, “FER-YOLO-Mamba: Facial expression detection and classification based on selective state space,” arXiv preprint arXiv:2405.01828 , 2024

    2024 arXiv

  2. [7]

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

    A. Gu and T. Dao, “Mamba: Linear-time sequence modeling with selective state spaces,” arXiv preprint arXiv:2312.00752 , 2023

    2023 arXiv

  3. [8]

    CSRNet: Dilated convolutional neural networks for understanding the highly congested scenes,

    Y . Li, X. Zhang, and D. Chen, “CSRNet: Dilated convolutional neural networks for understanding the highly congested scenes,” in 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2018, pp. 1091–1100

    2018

  4. [9]

    Learning from synthetic data for crowd counting in the wild,

    Q. Wang, J. Gao, W. Lin, and Y . Yuan, “Learning from synthetic data for crowd counting in the wild,” in 2019 IEEE/CVF Con- ference on Computer Vision and Pattern Recognition (CVPR) , 2019, pp. 8190–8199

    2019

  5. [13]

    Rotated multi-scale interaction network for referring remote sensing image segmentation,

    S. Liu, Y . Ma, X. Zhang, H. Wang, J. Ji, X. Sun, and R. Ji, “Rotated multi-scale interaction network for referring remote sensing image segmentation,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2024, pp. 26 658–26 668

    2024

  6. [14]

    MambaIRv2: Attentive state space restoration,

    H. Guo, Y . Guo, Y . Zha, Y . Zhang, W. Li, T. Dai, S.-T. Xia, and Y . Li, “MambaIRv2: Attentive state space restoration,” arXiv preprint arXiv:2411.15269, 2024

    2024 arXiv

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.