{"id":"0ada79a8-7125-437a-b9b7-42ae3b80a7f0","arxiv_id":"2505.23214","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"SAMamba, combining a frozen SAM2/Hiera encoder with Vision Mamba blocks and three lightweight modules, achieves state-of-the-art infrared small target detection on NUAA-SIRST, IRSTD-1k, and NUDT-SIRST.","lead":"This paper presents SAMamba, an infrared small-target detection network that pairs a pre-trained SAM2 encoder with Mamba-style sequence modeling and three new modules for domain adaptation, global context, and multi-scale fusion. It reports state-of-the-art scores on three public infrared small-target benchmarks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The SOTA claim hinges on Table 1, where SAMamba's 1024x1024 training protocol and combined loss are not shown to be matched by the baselines; the +1% margins over ISNet and HCFNet may be protocol artifacts.","rationale":"The reader's weakest_assumption correctly identifies the most load-bearing threat to the paper's strongest claim. The empirical superiority statement is meaningful only if the comparison is fair, and the paper provides no evidence that baselines were trained under SAMamba's protocol. In fact, Section 4.1's description of SAMamba's pipeline and Table 4's 1024x1024 FLOPs calculation suggest the authors may have computed baseline numbers at SAMamba's resolution without retraining. This is especially significant because the reported margins over the closest competitors are about 1 IoU point, while the paper's own resolution ablation (Figure 6a) shows that resolution changes can be several points. A second, smaller concern is that DPCF's 'spatially adaptive' gate is defined by a single scalar alpha (Eq. 13) expanded to all positions, so the module is not spatially adaptive as claimed; this bears on the novelty/ablation interpretation but does not directly falsify the empirical SOTA claim. I therefore agree with the reader's CONDITIONAL verdict: the central claim stands only if the baseline comparison is shown to be protocol-matched, which is exactly the concrete test proposed.","tokens_in":15222,"tokens_out":4319,"duration_ms":46597,"concrete_test":"Use the official releases of ISNet (CVPR'22) and HCFNet (ICME'24); retrain each on the identical NUAA-SIRST train/val split used for SAMamba, with SAMamba's full protocol (1024x1024 dynamic crop, Adam lr 1e-4, 300 epochs, batch 2, combined SoftIoU+Dice+Focal loss). Recompute their IoU/nIoU/F1 in Table 1. If ISNet or HCFNet reaches 81.08% IoU or the margin drops below roughly 0.5%, the reported SOTA advantage is a protocol artifact; if both remain clearly below, the central claim survives this check.","verdict_should_be":"UNCHANGED","load_bearing_attack":"In Section 4.1 the authors state that SAMamba is trained on 1024x1024 inputs with dynamic scaling/cropping, Adam (lr 1e-4), 300 epochs, batch 2, and the combined SoftIoU+Dice+Focal loss. Section 4.3 draws the central claim from Table 1: SAMamba exceeds ISNet by +1.06 IoU and HCFNet by +0.99 IoU on NUAA-SIRST. The paper never states that any baseline was retrained under this protocol. The original ISNet, UIUNet, and HCFNet papers use different resolutions (commonly 512x512) and often different loss functions and schedules. Figure 6(a) shows that resolution alone changes IoU by several points, so a 1-point margin could be explained by resolution/training recipe. Table 4 additionally reports baseline FLOPs at 1024x1024 without evidence that the baselines were actually configured and run at that resolution. If the Table 1 numbers for prior methods are copied from their original papers, the comparison is not apples-to-apples, and the 'consistently achieves state-of-the-art' claim is not yet established. This is the load-bearing assumption: the head-to-head must be fair, and the paper does not demonstrate it.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SAMamba, an infrared small target detection (ISTD) architecture that combines a frozen SAM2 Hiera-S backbone with a new Feature Selection Adapter (FS-Adapter), a Cross-Channel State-Space Interaction (CSI) module based on Vision Mamba in the skip connections, and a Detail-Preserving Contextual Fusion (DPCF) module in the decoder. The method is evaluated on three ISTD benchmarks (NUAA-SIRST, IRSTD-1k, NUDT-SIRST) with IoU, nIoU, and F1 metrics. The authors report state-of-the-art results on all three datasets, together with component ablations, hyperparameter sensitivity studies, and a computational analysis. The central empirical claim is that SAMamba consistently outperforms prior methods, with margins of roughly 1 IoU point over ISNet and HCFNet on NUAA-SIRST and larger margins on IRSTD-1k and NUDT-SIRST.","tokens_in":15447,"tokens_out":5722,"duration_ms":53817,"significance":"If the reported comparisons are fair, SAMamba would be a credible new state of the art for ISTD on the three chosen benchmarks. The paper has several concrete strengths: it releases code, performs systematic ablations that isolate each module's contribution with monotonic gains, includes sensitivity analyses for internal hyperparameters, and provides an explicit computational analysis. The architectural combination of SAM2's hierarchical features with Mamba-based sequence modeling is sensible and timely. However, the significance of the central claim is currently weakened by incomplete evidence about the fairness of the baseline comparisons, by a formal inconsistency in the description of the DPCF gating mechanism, and by the absence of any statistical validation. These issues do not invalidate the general approach, but they need to be resolved before the state-of-the-art claim can be accepted.","major_comments":[{"comment":"The central claim in §4.3 that SAMamba 'consistently achieves state-of-the-art performance across all three datasets and all three evaluation metrics' rests on Table 1. SAMamba is trained at 1024×1024 resolution with dynamic scaling/cropping, Adam (lr 1e-4), 300 epochs, batch size 2, and a combined SoftIoU+Dice+Focal loss. The paper does not state whether any baseline was retrained under this protocol, and the original ISNet, UIUNet, and HCFNet papers use different resolutions (commonly 512×512) and different training recipes. Figure 6(a) shows that resolution alone can change IoU by several points, so the ~1-point margins on NUAA-SIRST (+1.06 over ISNet, +0.99 over HCFNet) may reflect protocol differences rather than architectural superiority. Please specify the exact source of every baseline number in Table 1 (original publication vs. reimplementation), and if numbers are copied, retrain the main baselines under the same protocol or report both original and matched-protocol results. The same concern applies to Table 4, where baseline FLOPs are reported for a 1024×1024 input without evidence that the baselines were configured and run at that resolution.","section":"§4.1 and Table 1"},{"comment":"The DPCF module is described as a 'learnable, spatially-adaptive gating mechanism,' but the formal description does not support that claim. A single learnable parameter α' ∈ R^{1×1×1} is introduced and then expanded spatially and channel-wise to α ∈ R^{H×W×C/4}. Expanding a scalar produces a tensor in which all entries are identical, so β = sigmoid(α) is a single scalar value, not a per-location per-channel gate. This contradicts the subsequent statement that the network learns weights 'for each spatial location and channel group.' Please clarify the actual parameterization: either α is a learned tensor with spatial and channel dimensions, in which case the text should say so and Table 4's parameter counts should be updated accordingly, or the gating is a scalar fusion weight, in which case the 'spatially-adaptive' language and the ablation claims in Table 3(c) need to be revised.","section":"§3.5, Eqs. (13)-(14)"},{"comment":"The dimensions in the FS-Adapter token-selection step are inconsistent. F'_t ∈ R^{HW×C} and ξ ∈ R^C, but t_i is written as belonging to R^{HW} while the index i is ranged to C. If t_i is a token vector of length C, the index should run over HW and the notation should be changed accordingly; if t_i is a per-channel spatial map, the cosine similarity sim(t_i, ξ) is undefined because the two arguments have different dimensions. The intended mechanism is recoverable, but the formal description is not self-consistent and should be corrected before the module can be implemented or evaluated from the paper alone.","section":"§3.3, Eqs. (6)-(7)"},{"comment":"All main results and ablations are reported as point estimates with no error bars, multiple seeds, or statistical significance tests. The headline margins over ISNet and HCFNet on NUAA-SIRST are only about 1 IoU point, and Figure 6(a) shows that experimental conditions such as input resolution can shift performance by several points. Please report the mean and standard deviation over at least three seeds for the main comparison in Table 1 and for the component ablation in Table 2, and state whether the observed component-wise gains are stable across seeds. This is necessary to support the word 'consistent' in the central claim.","section":"§4.4 and Table 1"}],"minor_comments":[{"comment":"The variable names in Eq. (16) are reversed relative to standard convention: the text says 'ŷ represents the ground truth, y is the predicted mask,' but the usual notation is that y is the ground truth and ŷ is the prediction. Please swap the definitions or the symbols for clarity.","section":"§3.6, Eq. (16)"},{"comment":"The UIUNet reference appears twice, once as Wu et al. (2022) and once as Wu et al. (2023), for what appears to be the same paper. Please consolidate the duplicate entry.","section":"References"},{"comment":"The dagger symbol (†) next to ISNet in Table 4 is not explained in the table caption or the text. Please add a footnote describing what it indicates.","section":"Table 4"},{"comment":"The sentence explaining why Hiera-L degrades performance says this is 'likely due to overfitting on the relatively small ISTD datasets,' but no evidence such as training/validation curves or regularization settings is provided. Please either qualify this as a hypothesis or support it with data.","section":"§4.4, Figure 6(c)"}],"recommendation":"major_revision","confidential_remarks":"The paper is an empirical architecture study with no evidence of circular reasoning; the self-citation of HCFNet is normal practice. The main risk to the paper's central claim is comparison fairness, which can be addressed with a careful reimplementation or transparent sourcing of baseline numbers. The DPCF gating inconsistency is also fixable in revision. I see no grounds for rejection, but the current manuscript should not be accepted without the requested clarifications."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"SAMamba is a serious, workmanlike entry in the infrared small target detection line. The new bit is concrete: a frozen SAM2/Hiera encoder with adapter modules, Vision Mamba blocks inside the skip connections, and a gated fusion in the decoder. The three named modules (FS-Adapter, CSI, DPCF) are not in prior work, and the ablation shows each one adds a meaningful, monotonic chunk of IoU on NUAA-SIRST. That is real evidence the architecture, not just the backbone, is carrying weight. The paper also ships code and states its limitations clearly, including failure cases on complex backgrounds and low SCR.\n\nThe central SOTA claim is plausible but not nailed down. The paper never says whether the baselines in Table 1 were retrained under SAMamba's 1024x1024, dynamic-crop, Adam, SoftIoU+Dice+Focal protocol. If those numbers are copied from the original papers (the common practice in this subfield), then the +1.0 IoU margin over ISNet and HCFNet on NUAA-SIRST could be largely a protocol artifact — Figure 6(a) shows resolution alone moves SAMamba's IoU by several points, so a one-point margin is within that range. This is the load-bearing soft spot, and it needs a sentence or a supplementary table to fix.\n\nThere are smaller issues worth naming. The DPCF gate is described as spatially adaptive, but alpha is a single scalar per stage that gets broadcast — the \"spatial\" language oversells it. The hyperparameter sensitivity study (Table 3) is run on the NUAA-SIRST test set, which is the same benchmark where the headline numbers are reported; that is a mild selection-on-test smell, common in this literature but still worth a flag. There are no error bars or significance tests, and the FS-Adapter text is a bit tangled — equation (6) is written as token selection but the prose says channel re-weighting.\n\nNone of this sinks the paper. The architecture is sensible, the gains are consistent across three datasets, and the ablations back the components. It deserves a serious referee, and I'd want the baseline-protocol question resolved before trusting the SOTA claim. If the baselines were actually retrained under the same protocol, this is a clean, useful result. If not, it is still a decent architecture but the margin over ISNet/HCFNet is unproven.\n\nRecommended: send to peer review; ask for baseline protocol details, error bars, and a de-overclaimed description of the DPCF gate.","headline":"Plausible new SOTA for infrared small target detection, but the comparison protocol against baselines is underspecified and must be checked.","tokens_in":16051,"tokens_out":2387,"would_cite":false,"duration_ms":24111,"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":"SAMamba claims state-of-the-art infrared small-target detection on three public benchmarks by pairing SAM2's frozen hierarchical encoder with Mamba's selective state-space modeling.","keywords":["infrared small target detection","segment anything model","vision mamba","state space models","domain adaptation","feature fusion","U-Net","small object detection"],"falsifier":"Retrain ISNet, HCFNet, and UIUNet with SAMamba's exact training recipe, including 1024x1024 dynamic crops, Adam, and SoftIoU plus Dice plus Focal loss for 300 epochs, then compare on NUAA-SIRST; if their IoU climbs to within about one point of SAMamba's 81.08, the claimed architectural advantage is a protocol artifact, whereas a persistent gap of several points would support the central claim.","tokens_in":15003,"feed_emoji":"🎯","tokens_out":6574,"duration_ms":59060,"temperature":0.7,"pith_summary":"The paper sets out to show that combining SAM2's hierarchical feature learning with Mamba's selective state-space sequence modeling yields a detector that outperforms existing CNN-based methods for infrared small target detection, where targets occupy under 0.15% of the image and blend into clutter. The authors claim this is achieved by freezing the SAM2 encoder and adding three lightweight modules that adapt features to the infrared domain, model global context with linear complexity, and preserve small-target detail during multi-scale fusion. Across NUAA-SIRST, IRSTD-1k, and NUDT-SIRST, the reported IoU, nIoU, and F1 scores exceed all compared methods. If the comparisons hold up under matched training protocols, the result matters because it gives a straightforward recipe for adapting a large vision foundation model to a sparse, low-contrast detection task without full fine-tuning.","feed_headline":"SAMamba tops three infrared small-target detection benchmarks","feed_subtitle":"Reports 81.08 IoU on NUAA-SIRST and 93.13 IoU on NUDT-SIRST, beating prior CNN detectors.","key_machinery":"The load-bearing machinery is a three-part adaptation of a frozen SAM2 encoder. The FS-Adapter re-weights tokens by cosine similarity to a learnable task embedding and then mixes channels through a learned matrix, inserted before each Hiera block to close the natural-to-infrared gap. The CSI module splits skip-connection features into four channel groups, runs each through a Vision Mamba state-space block, recombines same-index channels across groups, and applies channel and spatial attention, giving global context at linear cost. The DPCF module upsamples deep features and uses a single learnable scalar expanded into spatial-channel gates to fuse high- and low-resolution segments, so the network can choose detail versus context per location.","core_discovery":"The paper's central claim is that SAMamba, a U-Net with a frozen SAM2 Hiera-S encoder and three lightweight modules, achieves the best published results on NUAA-SIRST, IRSTD-1k, and NUDT-SIRST across IoU, nIoU, and F1. On NUAA-SIRST it reports 81.08% IoU, 79.17% nIoU, and 89.55% F1; on IRSTD-1k 73.53% IoU, 68.99% nIoU, and 84.75% F1; on NUDT-SIRST 93.13% IoU, 93.15% nIoU, and 96.44% F1. The authors attribute the gains to domain adaptation via the FS-Adapter, long-range context via the CSI module's Mamba-based state-space interaction, and detail preservation via DPCF's adaptive gating fusion.","pith_inferences":["Editorial inference: if the protocol-matched comparison confirms the margins, the FS-Adapter design could transfer directly to other low-contrast or non-visible-spectrum tasks, such as radar or hyperspectral segmentation, where only small labeled sets exist.","Editorial inference: the paper's own limitation analysis suggests the next testable extension is video, where feeding temporal frames through the same CSI state-space blocks could exploit motion cues, a direction the authors mention as future work.","Editorial inference: because DPCF uses a single scalar expanded to gates, replacing it with per-channel learned gates might further improve fusion on heterogeneous backgrounds; this is a cheap experiment the paper does not run.","Editorial inference: the large gain on NUDT-SIRST (93.13 IoU) relative to CNN baselines may reflect the synthetic dataset's clean target shapes, so a real-world low-SCR benchmark would clarify whether the architecture's advantage persists under sensor noise."],"forward_implications":["A frozen SAM2 Hiera-S encoder, adapted by lightweight modules, is sufficient to beat fully trained CNN detectors on three ISTD benchmarks.","The Mamba-based CSI module's linear-complexity global context modeling is a practical alternative to attention for high-resolution infrared inputs.","The component ablations attribute the improvement to each module: +0.89 IoU from FS-Adapter, +2.47 from CSI, and +2.28 from DPCF over the Hiera-based baseline.","At 1024x1024 input, SAMamba's 37.18M-parameter Hiera-S variant runs at 6.39 FPS on an RTX 3090, comparable to ISNet and HCFNet while reporting higher accuracy.","Domain adaptation through token-level selection plus channel mixing can be reused in other tasks that use a frozen foundation-model encoder on a different imaging modality."],"supporting_citations":[{"why":"Supplies the SAM2 Hiera backbone whose frozen features SAMamba adapts.","marker":"Ravi et al., 2024"},{"why":"Supplies the Vision Mamba block used inside the CSI module for bidirectional state-space sequence modeling.","marker":"Zhu et al., 2024"},{"why":"Introduces selective state-space models that give the CSI module linear-complexity global context.","marker":"Gu and Dao, 2023"},{"why":"Defines the Hiera hierarchical encoder architecture that SAM2 uses to provide multi-scale features.","marker":"Ryali et al., 2023"},{"why":"Provides the NUAA-SIRST benchmark, the ACM baseline, and the nIoU metric used in evaluation.","marker":"Dai et al., 2021a"},{"why":"Provides the ISNet baseline, a close competitor on NUAA-SIRST, and its edge-aware shape modeling approach.","marker":"Zhang et al., 2022b"},{"why":"Provides the HCFNet baseline and hierarchical context fusion method that SAMamba compares against.","marker":"Xu et al., 2024"},{"why":"Provides the UIUNet baseline, a dual U-Net competitor evaluated on all three datasets.","marker":"Wu et al., 2022"},{"why":"Provides the IRSTD-1k dataset, the diverse-scenario benchmark used for evaluation.","marker":"Zhang et al., 2022c"},{"why":"Provides the NUDT-SIRST dataset, the synthetic small-target benchmark with extremely small and dim targets.","marker":"Li et al., 2022"}],"fun_headline_variants":["SAMamba beats prior SOTA on three infrared small-target datasets","SAMamba fuses SAM2 and Mamba to outdo prior IR small-target detectors","SAMamba: hierarchical vision plus selective state space for IR small targets","SAMamba tops three IR small-target benchmarks with novel state-space fusion","SAMamba: state-space model sets new records on infrared small-target detection"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the baseline methods in Table 1 were trained and evaluated under the same protocol as SAMamba, including 1024x1024 resolution, dynamic cropping, Adam optimization, and the combined SoftIoU plus Dice plus Focal loss; if the published baseline numbers used a different protocol, the reported margins could reflect training conditions rather than the architecture.","fun_headline_variants_meta":{"raw":{"variants":["SAMamba beats prior SOTA on three infrared small-target datasets","SAMamba fuses SAM2 and Mamba to outdo prior IR small-target detectors","SAMamba: hierarchical vision plus selective state space for IR small targets","SAMamba tops three IR small-target benchmarks with novel state-space fusion","SAMamba: state-space model sets new records on infrared small-target detection"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001329,"raw_usage":{"total_tokens":5445,"prompt_tokens":1020,"completion_tokens":4425,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":636,"completion_tokens_details":{"reasoning_tokens":4330}},"tokens_in":636,"tokens_out":4425,"duration_ms":33711,"temperature":1.0,"reasoning_tokens":4330,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T12:51:16.362103+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain ISNet, HCFNet, and UIUNet with SAMamba's exact training recipe, including 1024x1024 dynamic crops, Adam, and SoftIoU plus Dice plus Focal loss for 300 epochs, then compare on NUAA-SIRST; if their IoU climbs to within about one point of SAMamba's 81.08, the claimed architectural advantage is a protocol artifact, whereas a persistent gap of several points would support the central claim.","supporting_citations":[{"cited_title":", author Hu, Y.T","cited_arxiv_id":null,"evidence_quote":"Defines the Hiera hierarchical encoder architecture that SAM2 uses to provide multi-scale features."}],"review_version":1}