{"id":"a1e809f9-4548-47bf-a6aa-7db58f701c16","arxiv_id":"2501.06697","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"Mamba-MOC, a Mamba-based counting network with a cross-scale interaction module and a context state space model, reports state-of-the-art MSE and WMSE on NWPU-MOC.","lead":"This paper applies an efficient Mamba neural network architecture to count different object types in aerial images, reporting better accuracy than five earlier methods on a standard benchmark. The practical payoff is a possible speed-versus-accuracy improvement for remote sensing inventories such as counting ships, buildings, and trees.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Central SOTA claim rests on unverified baseline comparability; Table I also contains arithmetic inconsistencies (SFCN and MCC overall MSEs), so the headline result needs a controlled reproduction before it can be accepted.","rationale":"I read the paper as making an empirical state-of-the-art claim, not a mathematical guarantee. The load-bearing condition for that claim is a controlled comparison: identical density-map generation, identical split, and identical evaluation protocol across all methods. The paper does not establish this, and the latest baselines predate the dataset, making the provenance of their numbers essential. The arithmetic inconsistencies in Table I (SFCN and MCC overall MSEs not matching the stated inter-category average) strengthen the reader's concern by showing the table cannot be taken at face value. I do not see an internal contradiction that would overturn the result if the reported numbers are correct; the architecture is plausible, the ablation is consistent, and the method outperforms all listed baselines. But the lack of code, error bars, and baseline protocol details warrants a conditional verdict pending reproduction. The reader's verdict is therefore appropriate, and this stress-test does not change it.","tokens_in":5718,"tokens_out":8338,"duration_ms":78000,"concrete_test":"Fetch the official NWPU-MOC benchmark code and density-map generation function from [12]; retrain or execute the five baselines (CSRNet, SFCN, PSGCNet, DSACA, MCC) with the official train/test split and the paper's stated Gaussian sigma=4, kernel size 15, 512x512 inputs, and compare reproduced MSE/WMSE against Table I. If any baseline moves by more than about 0.5 in MSE, the comparison is not controlled. Also recompute overall MSE as the mean of the six per-category MSEs for SFCN and MCC; if the values do not equal 12.1374 and 11.1826, respectively, Table I needs correction before it can support a state-of-the-art claim.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim — state-of-the-art MSE 9.5794 / WMSE 27.2012 on NWPU-MOC — is only as strong as the five baseline entries in Table I. The paper specifies its own density-map protocol (Gaussian kernel, bandwidth 4, size 15; Section IV-A) and its own train/test split, but never states whether the baseline numbers from [8]–[12] were produced under the identical protocol. This matters because [10] PSGCNet (2022) and [11] DSACA (2021) predate the NWPU-MOC dataset (2024), so their entries cannot come from the original papers; they must come from a reimplementation or a third-party benchmark whose training details are undisclosed. The concern is reinforced by internal inconsistencies in the same table: for SFCN the reported overall MSE (14.8041) is not the average of the six listed category MSEs (12.1374), and for MCC the reported 11.4826 differs from the computed 11.1826. These discrepancies do not by themselves overturn the claimed wins, but they make the headline result conditional on a controlled, code-level reproduction rather than the table as printed.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","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.","tokens_in":6029,"tokens_out":9795,"duration_ms":80042,"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":[{"comment":"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.","section":"Section IV-A and Table I"},{"comment":"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.","section":"Table I"},{"comment":"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.","section":"Section III-C, Eq. (8)"},{"comment":"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.","section":"Section IV-C, Table II"}],"minor_comments":[{"comment":"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].","section":"Section IV-A"},{"comment":"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.","section":"Section IV-B and Table I"},{"comment":"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).","section":"Section II, Eq. (1)"},{"comment":"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.","section":"Section III-C"},{"comment":"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.","section":"Section III-C"}],"recommendation":"major_revision","confidential_remarks":"The manuscript has a plausible core idea and the reported improvements are internally consistent in the ablation, but the main comparison table is not yet trustworthy. The arithmetic discrepancies and the unverified baseline protocol are likely fixable in revision, and I would like to see the corrected table and a reproducibility statement before judging the SOTA claim."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague, quick take on Mamba-MOC. The genuinely new bit is applying a Mamba backbone to multi-category remote sensing counting and the small CSSM twist on the SSM output equation (adding a local query to C and a local feature to the output). That is a real, if modest, extension. The cross-scale interaction is standard FPN-style fusion with a VSS block in the middle, not much more.\n\nWhat the paper does well: the ablation is internally consistent. Baseline (backbone+FPN) gives 10.38 MSE; adding CIM drops it to 10.35; adding CSSM drops it to 9.58. If those numbers are reproducible, the components do something. The idea of using a local convolution to compensate for causal scanning in 2D Mamba is reasonable and follows directly from MambaIRv2.\n\nThe soft spots are not minor. First, Table I has arithmetic errors: for SFCN the category MSEs average to 12.14, not the 14.80 printed; for MCC the average is 11.18, not 11.48. That makes me distrust the whole table until corrected. Second, the comparison protocol is unverified. PSGCNet and DSACA predate NWPU-MOC, so their entries must come from a reimplementation, but the paper never states whether the density-map generation, kernel size, or split were identical to theirs. The paper gives its own kernel settings but doesn't confirm the baselines were run under the same conditions. That is a load-bearing gap. Third, no code, no error bars, no runtime comparison.\n\nThe central claim -- 'we are SOTA' -- is therefore conditional. The method is plausible, and it's likely Mamba can be competitive for this task, but the numbers as printed don't support the headline. This is not a case where the whole enterprise is wrong; it's a case where the paper needs a controlled reproduction and a corrected table before I'd trust the exact figures.\n\nWho gets value: practitioners working on remote sensing counting who want a Mamba baseline to compare against, and people tracking state-space models in dense prediction. Not a conceptual breakthrough.\n\nRecommendation: send it to peer review, but require the authors to fix the table, disclose or re-run the baselines under identical protocol, and ideally release code or error bars. With that, it could be a solid workshop-to-conference level paper.","headline":"Plausible Mamba-based counting network, but the SOTA claim rests on a Table I with arithmetic errors and unverified baseline settings.","tokens_in":6493,"tokens_out":3522,"would_cite":false,"duration_ms":31814,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"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.","keywords":["remote sensing","object counting","Mamba","state-space model","multicategory counting","density map estimation","cross-scale interaction","NWPU-MOC"],"falsifier":"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.","tokens_in":5536,"feed_emoji":"🛰️","tokens_out":7822,"duration_ms":63286,"temperature":0.7,"pith_summary":"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.","feed_headline":"Mamba model sets benchmark records for remote object counting","feed_subtitle":"Mamba-MOC cuts MSE to 9.58 and WMSE to 27.20, winning five of six object categories.","key_machinery":"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.","core_discovery":"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.","pith_inferences":["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."],"forward_implications":["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."],"supporting_citations":[{"why":"Supplies the Vmamba backbone used to extract the hierarchical features in Mamba-MOC.","marker":"[5]"},{"why":"Defines the selective structured state-space model whose discretized dynamics the context SSM modifies.","marker":"[7]"},{"why":"Provides the NWPU-MOC dataset, its six-category grouping, and the MSE and WMSE evaluation metrics.","marker":"[12]"},{"why":"Motivates treating the output matrix C as a query and injecting local context into the state-space output.","marker":"[14]"},{"why":"Supplies the gating mechanism that fuses cross-scale features in the cross-scale interaction module.","marker":"[13]"},{"why":"Baseline CNN counting method whose per-category results are compared in the main table.","marker":"[8]"},{"why":"Baseline crowd-counting method used as a comparison in the remote counting benchmark.","marker":"[9]"},{"why":"Baseline remote-sensing dense counting network used as a comparison.","marker":"[10]"},{"why":"Baseline multi-class object counting convnet used as a comparison.","marker":"[11]"}],"fun_headline_variants":["Mamba-MOC: first Mamba-based remote object counter","Mamba-MOC cuts remote counting MSE to 9.58","Linear-complexity Mamba beats CNN and Transformer counters","Mamba-MOC: cross-scale interaction for remote counting","Mamba reimagined for multicategory remote object counting"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"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.","fun_headline_variants_meta":{"raw":{"variants":["Mamba-MOC: first Mamba-based remote object counter","Mamba-MOC cuts remote counting MSE to 9.58","Linear-complexity Mamba beats CNN and Transformer counters","Mamba-MOC: cross-scale interaction for remote counting","Mamba reimagined for multicategory remote object counting"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000618,"raw_usage":{"total_tokens":2842,"prompt_tokens":892,"completion_tokens":1950,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":508,"completion_tokens_details":{"reasoning_tokens":1866}},"tokens_in":508,"tokens_out":1950,"duration_ms":14434,"temperature":1.0,"reasoning_tokens":1866,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T20:54:48.602537+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"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.","supporting_citations":[{"cited_title":"NWPU-MOC: A benchmark for fine-grained multicategory object counting in aerial images,","cited_arxiv_id":null,"evidence_quote":"Provides the NWPU-MOC dataset, its six-category grouping, and the MSE and WMSE evaluation metrics."},{"cited_title":"Rotated multi-scale interaction network for referring remote sensing image segmentation,","cited_arxiv_id":null,"evidence_quote":"Supplies the gating mechanism that fuses cross-scale features in the cross-scale interaction module."},{"cited_title":"CSRNet: Dilated convolutional neural networks for understanding the highly congested scenes,","cited_arxiv_id":null,"evidence_quote":"Baseline CNN counting method whose per-category results are compared in the main table."},{"cited_title":"Learning from synthetic data for crowd counting in the wild,","cited_arxiv_id":null,"evidence_quote":"Baseline crowd-counting method used as a comparison in the remote counting benchmark."},{"cited_title":"PSGCNet: A pyramidal scale and global context guided network for dense object counting in remote-sensing images,","cited_arxiv_id":null,"evidence_quote":"Baseline remote-sensing dense counting network used as a comparison."},{"cited_title":"Dilated- scale-aware category-attention convnet for multi-class object counting,","cited_arxiv_id":null,"evidence_quote":"Baseline multi-class object counting convnet used as a comparison."}],"review_version":1}