{"id":"4e414864-0d8f-423b-aa0a-4fa6fe5e28e6","arxiv_id":"2508.21135","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A Mamba-based fusion network with a channel-aware decoder reports competitive or state-of-the-art results on RGB-thermal and RGB-depth hidden-object detection benchmarks.","lead":"A computer-vision system fuses RGB photos with thermal or depth images, using a Mamba-based neural network, to detect hidden or camouflaged objects. It reports small gains on several benchmarks against existing fusion methods, but without code or error bars.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"ACOD-12K 'significantly outperforms' is contradicted by Table 4: Ours is tied with or only 0.007 above RISNet, within the paper's own admitted non-significance; no ablation isolates the fusion novelty.","rationale":"Good-faith reading: the paper proposes a plausible VMamba-based multimodal fusion architecture and provides a broad evaluation. The strongest claim has two parts: the architecture is a useful baseline, and it is state-of-the-art or competitive. The first part may well be true, but the second is the load-bearing empirical assertion. I find the most decisive problem in Table 4: the numbers as printed do not support 'significantly outperforms.' The abstract's full-precision values make this worse: Sα=0.865 vs RISNet's rounded 0.87; Eφ ties; only Fβ is higher by 0.007. The authors themselves disclaim statistical significance in the Discussion. Since no error bars or repeated runs are provided, the reader cannot distinguish the claimed advantage from noise. The reader's weakest_assumption focuses on cross-paper comparability of baselines (different backbones and resolutions); that is real and related, but the sharper issue is the internal inconsistency in the headline ACOD-12K result. I also note the absence of ablations separating the MMFF and decoder C-computation from Sigma; this matters because Sigma (VMamba-S) is nearly as strong on MFNet/PST900, so the marginal contribution of the proposed fusion arrangement is unknown. These concerns are not accusations of misconduct; they are requests for evidence. A conditional accept with a requirement to re-benchmark and report uncertainty remains appropriate, so the reader's verdict is retained.","tokens_in":15002,"tokens_out":5036,"duration_ms":50137,"concrete_test":"Re-run RISNet and HiddenObject on ACOD-12K under the paper's exact protocol (same resolution, optimizer, epochs, and, for a controlled comparison, VMamba-S backbone) with 5 random seeds and report full-precision Sα/Fβ/Eφ with 95% confidence intervals. If the Fβ difference (0.807 vs 0.800) is within seed noise, or if RISNet's Sα remains ≥0.865, the 'significantly outperforms' claim fails and the paper should be revised to 'competitive' with error bars.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central empirical claim—'significantly outperforms existing methods across all metrics' on ACOD-12K (Abstract, §4.2)—is contradicted by the paper's own Table 4. RISNet is listed at Sα=0.87, Fβ=0.80, Eφ=0.97; Ours is listed at Sα=0.87, Fβ=0.81, Eφ=0.97, with the abstract giving full precision 0.865/0.807/0.965. On Sα, ours is below RISNet (0.865 < 0.87 if RISNet's 0.87 is rounded from ≥0.865); on Eφ they are tied; the only favorable difference is Fβ=+0.007. The Discussion explicitly says 'the observed fluctuations are not statistically significant in most cases,' yet no significance test, error bar, or multi-seed result is reported. The SOTA claim therefore rests on a rounding-level difference and is internally inconsistent. A second contributor is that no ablation isolates the proposed MMFF arrangement or the decoder's C-from-higher-level computation from Sigma (whose VMamba-S baseline is the closest competitor in Table 2); without such an ablation, even the 'competitive' numbers cannot be attributed to the claimed novelty. This does not show the architecture is wrong, but it does mean the paper's strongest claim is not established by the evidence presented.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes HiddenObject, a multimodal fusion framework built on VMamba SS2D modules. It uses a dual-stream encoder with shared weights for RGB and X-modality, an MMFF module that scans a concatenated RGB-X sequence and adds a reverse-scan flip-back branch, and a channel-aware decoder in which the SS2D matrix C is computed from higher-level or RGB features while A, B, and Δ come from lower-level features. The method is evaluated on RGB-T semantic segmentation (MFNet, PST900), RGB-D segmentation (NYU Depth V2, SUN RGB-D), and RGB-D concealed-object detection (ACOD-12K). The authors report the highest mIoU on MFNet and PST900 and claim significant improvements over all existing methods on ACOD-12K.","tokens_in":15397,"tokens_out":3961,"duration_ms":37417,"significance":"If the comparisons were properly controlled, the MMFF arrangement and the decoder's asymmetric use of the RGB feature would be a useful baseline for modality-agnostic hidden object detection, especially given the limited number of Mamba-based multimodal fusion studies on concealed-object benchmarks. The paper also benefits from a broad benchmark suite spanning thermal and depth modalities. However, the current evidence does not establish the stated SOTA claims: baseline numbers are taken from heterogeneous settings, the ACOD-12K 'significant' claim conflicts with the paper's own table and discussion, and no ablation isolates the proposed components. The architecture may well be competitive, but the validation as presented is insufficient.","major_comments":[{"comment":"The claim that 'our model significantly outperforms existing methods across all metrics' on ACOD-12K is not supported by Table 4. Compared with RISNet, Ours is 0.865 vs 0.87 in Sα (lower after rounding), 0.807 vs 0.80 in Fβ (+0.007), and 0.965 vs 0.97 in Eφ (tied within rounding). No error bars or significance tests are reported, and the Discussion in §4.2 admits that 'the observed fluctuations are not statistically significant in most cases.' This is an internal inconsistency in the paper's strongest claim; it must be replaced by a defensible statement and supported by multi-seed runs or statistical tests.","section":"Abstract, §4.2, Table 4"},{"comment":"The SOTA and 'highest' claims rely on direct comparison with published baseline numbers obtained under different backbones, input resolutions, and training protocols. For example, Table 3 mixes ResNet-152, MiT-B3/B4, ViT-B, MobileNet-V2, and VMamba-S at resolutions from 480×480 to 640×640. Crucially, the text says Ours 'achieves the highest mIoU of 52.1 on SUN RGB-D,' but Table 3 lists Sigma (VMamba-S) at 52.4, so this is factually wrong. Likewise, on NYU, Ours is third (56.8) behind Sigma (57.0) and CMNeXt (56.9), not merely 'slightly behind CMNeXt.' The authors need to re-run baselines under their own protocol or clearly qualify the comparisons as cross-paper.","section":"§4.1, Tables 2-3"},{"comment":"No ablation isolates the novelty. The closest comparison, Sigma (VMamba-S), differs from Ours in multiple ways: fusion arrangement, decoder C computation, scaling parameters, shuffling, and upsampling. Without removing or replacing the proposed MMFF and decoder components one at a time, the reported gains cannot be attributed to the paper's contribution. A minimal ablation should include: (a) Sigma's fusion in the same encoder, (b) Ours without the reverse-scan flip-back addition, and (c) decoder C computed from the lower-level or RGB feature instead of the higher-level feature. This is necessary even to support 'competitive' rather than 'SOTA.'","section":"§3.3, §3.4, Tables 2-4"}],"minor_comments":[{"comment":"The caption says 'PSD900' but should be 'PST900.'","section":"Figure 6"},{"comment":"The caption uses 'ACOD-K12' while the rest of the paper uses 'ACOD-12K.'","section":"Figure 1"},{"comment":"The text says 'All methods were trained and evaluated on a cluster equipped with NVIDIA RTX 3090 GPUs.' This cannot be true for baseline scores copied from other papers; please clarify which numbers were re-run and which were taken from publications.","section":"§4.1 Implementation Detail"},{"comment":"The table would benefit from a column noting the evaluation protocol and training data, and the 'In. Size' column should be consistently reported for every method.","section":"Table 3"},{"comment":"The Discussion assigns all metric differences to 'not statistically significant' without reporting variances, confidence intervals, or seeds. Either provide the underlying statistics or remove the statistical phrasing.","section":"§4.2 Discussion"}],"recommendation":"major_revision","confidential_remarks":"The paper is in scope for a CV venue, but the internal contradictions on SUN RGB-D and ACOD-12K, together with the complete lack of ablations, make the current version unsupportable as-is. If the authors can re-run baselines under a consistent protocol, add component ablations, and soften the claims to match the evidence, the contribution could become publishable. I would also encourage the editor to require that the ACOD-12K 'significant' language be reconciled with the reported numbers."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know this paper before anything else: the architecture is a real, if incremental, extension of Sigma and CMX, but the headline SOTA claims on ACOD-12K are not supported by the paper's own table. It deserves a serious referee, but only with the expectation of major revision.\n\nWhat's actually new: the specific fusion design—reverse scan order, decoder computing the C matrix from the RGB feature, and scale-aware alignment before fusion—is clearly described and isn't in the cited prior work. The paper is straightforward about borrowing from CMX and Sigma, which I appreciate. The evaluation spans RGB-T and RGB-D datasets, and the qualitative results show plausible gains on small, occluded objects.\n\nThe soft spots are real and central. Table 4 lists RISNet at Sα=0.87 and Eφ=0.97, while the authors report 0.865 and 0.965. Calling that 'significantly outperforms' contradicts the paper's own Discussion, which admits 'the observed fluctuations are not statistically significant in most cases.' There are no error bars, no multi-seed runs, no significance tests. On top of that, baseline numbers come from papers with different backbones, input resolutions, and training protocols—nothing is re-run under the authors' setup. And there is no ablation that isolates the proposed fusion arrangement or the decoder's C-from-RGB computation from Sigma. So even the 'competitive' results can't be attributed to the claimed novelty.\n\nNone of this makes the architecture wrong. It's a plausible variation on an established program, and the numbers are in the right ballpark. But the paper overclaims, and the evidence for the specific novelties is missing.\n\nThis paper is for people working on Mamba-based fusion or hidden object detection; they'd get value from the design details and the cautionary tale about benchmark apples-to-oranges. I'd send it to peer review, because it's a genuine attempt with a clear method, but I'd condition acceptance on a unified re-benchmark, proper ablations, and toned-down claims.","headline":"A plausible Mamba fusion variant that is undermined by overclaimed SOTA numbers and missing significance tests; worth a careful revision, not a desk reject.","tokens_in":15831,"tokens_out":2371,"would_cite":false,"duration_ms":23555,"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":"HiddenObject claims a single Mamba-based fusion network outperforms prior RGB-thermal and RGB-depth detectors on four of five benchmarks.","keywords":["hidden object detection","camouflaged object detection","multimodal fusion","Mamba","state space models","RGB-thermal segmentation","RGB-depth segmentation","selective scan"],"falsifier":"Re-run every baseline in Tables 2-4 under HiddenObject's exact protocol: same optimizer, learning rate, epochs, batch size, 640x480 input, and same train/val splits, then check whether HiddenObject still holds the top mIoU, S_alpha, F_beta, and E_phi values. Separately, drop the thermal or depth channel at inference and measure the performance drop to test whether the claimed modality-agnostic fusion actually preserves detection when a sensor fails.","tokens_in":1825,"feed_emoji":"🔍","tokens_out":2240,"duration_ms":63801,"temperature":0.7,"pith_summary":"This paper tries to establish that one fusion architecture, HiddenObject, can detect hidden or partially concealed objects equally well whether the second sensor is thermal or depth, without redesigning the network per modality. The architecture pairs a shared-weight Mamba encoder with a multiscale fusion module and a decoder that lets the higher-level RGB feature control how the hidden state is read out. On the benchmarks tested, the paper reports the highest mean IoU on MFNet (61.5), PST900 (88.5), and SUN RGB-D (52.1), and best camouflaged-object-detection scores on ACOD-12K (S_alpha 0.865, F_beta 0.807, E_phi 0.965). The reason to care is practical: if true, one model and one training recipe cover RGB-thermal and RGB-depth concealed-object detection, improving on dedicated per-modality designs.","feed_headline":"Mamba fusion tops hidden-object detection on four benchmarks","feed_subtitle":"A shared encoder fuses RGB with thermal or depth, reaching 88.5 mIoU on PST900 and best scores on ACOD-12K.","key_machinery":"The central mechanism is the multimodal selective-scan fusion block (MMFF) plus a decoder-side assignment of the state-space readout. In a selective state-space scan, the output is y = C h, and which matrix C is chosen determines what gets extracted from the hidden state. The decoder computes A, B, and Delta from the lower-level feature, feeds the lower-level embedding into the scan, but computes C from the higher-level (RGB-anchored) feature, so the readout is conditioned on the richer modality. In MMFF, two modality features are linearly projected, scanned through SS2D in four directions, and the concatenated sequence is also scanned in reverse, flipped back, added, and re-split, then scal","core_discovery":"In the paper's own terms, the discovery is that a Mamba (selective state-space) backbone can serve as a modality-agnostic fusion engine for hidden object detection. The network encodes RGB and an X-modality (thermal or depth) with two weight-shared Vision Mamba branches, then fuses at four scales through a Multimodal Feature Fusion module whose cross-scan and reverse-scan sequence exchange lets each modality read the other's context. Its channel-aware decoder performs SS2D scans where the A, B, and Delta matrices come from lower-level features while matrix C, the readout from the hidden state, comes from the higher-level feature, giving the stronger modality control over the output. Reported","pith_inferences":["A fair test of the claimed margins would re-run each baseline in Tables 2-4 under HiddenObject's exact training protocol; if those numbers shift, the architecture claim may survive while the precise margins change.","The paper's modality-agnostic claim predicts that NIR, event-camera, or any aligned extra modality could be fed through the same X-modality branch with no architectural change, and that this should generalize; this is directly testable.","The decoder's choice to compute C from the higher-level feature rather than from the fused feature is not ablated; isolating that choice would show where the gain actually comes from.","Because the thermal and depth experiments share weights, the model could be trained jointly on RGB-T and RGB-D data in one pass; the paper does not test this combined setting."],"forward_implications":["One training protocol covers RGB-thermal, RGB-depth, and any future X-modality, so deploying the model in agriculture or robotics with different sensors does not require architecture changes.","The multiscale fusion plus Mamba linear-complexity design suggests the gains do not rely on a heavy transformer decoder; the VMamba-S backbone keeps the model comparatively light.","The decoder's C-from-higher-feature readout provides a natural handle for missing-modality scenarios: when one sensor drops out, the remaining modality can still drive the hidden-state readout.","The ACOD-12K results bring camouflaged-object-detection metrics into multimodal benchmarks, allowing structural measures such as S_alpha and E_phi rather than only mIoU.","Because the same code sets new bests across datasets, existing unimodal detectors that fail under occlusion and camouflage could be upgraded by swapping in the fusion mechanism without task-specific redesign."],"supporting_citations":[{"why":"Supplies the selective state-space (Mamba/S6) scan that the encoder, fusion module, and decoder all build on.","marker":"[27]"},{"why":"Defines Vision Mamba and the SS2D four-direction cross-scan/cross-merge used by the dual-stream encoder and decoder.","marker":"[62]"},{"why":"Sigma is the closest prior Mamba-based multimodal segmentation baseline on MFNet/PST900 with the VMamba-S backbone, and the fusion module is inspired by it.","marker":"[33]"},{"why":"CMX is the source of the multi-scale cross-modal fusion idea and supplies the AdamW training recipe adopted in the implementation.","marker":"[60]"},{"why":"MFNet provides one of the RGB-thermal datasets and the early-fusion baseline that the paper must beat.","marker":"[12]"},{"why":"PST900 provides the second RGB-thermal benchmark and the PSTNet baseline.","marker":"[14]"},{"why":"RISNet and the ACOD-12K benchmark provide the concealed-crop RGB-D dataset and the strongest published baseline on it.","marker":"[8]"},{"why":"CMNeXt is the strongest RGB-D semantic segmentation baseline on NYU Depth V2 and SUN RGB-D, including the MiT-B4 model that narrowly beats the paper on NYU.","marker":"[77]"}],"fun_headline_variants":["Mamba fusion beats unimodal and naive fusion on hidden objects","Weight-shared Mamba fuses RGB and thermal or depth for hidden objects","Mamba-based fusion tops hidden object detection across four benchmarks","Mamba backbone unifies multimodal cues for concealed object detection","Hidden object detection: Mamba fusion outperforms prior methods"],"cache_read_input_tokens":17536,"weakest_assumption_plain":"The reported superiority assumes that published baseline numbers from different backbones, input resolutions, and training recipes can be compared directly with a single new run under the authors' setup; if those baselines were re-run under the same protocol and scored differently, the claimed margins would change.","fun_headline_variants_meta":{"raw":{"variants":["Mamba fusion beats unimodal and naive fusion on hidden objects","Weight-shared Mamba fuses RGB and thermal or depth for hidden objects","Mamba-based fusion tops hidden object detection across four benchmarks","Mamba backbone unifies multimodal cues for concealed object detection","Hidden object detection: Mamba fusion outperforms prior methods"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000829,"raw_usage":{"total_tokens":3443,"prompt_tokens":717,"completion_tokens":2726,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":461,"completion_tokens_details":{"reasoning_tokens":2640}},"tokens_in":461,"tokens_out":2726,"duration_ms":15388,"temperature":1.0,"reasoning_tokens":2640,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T14:33:18.084016+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run every baseline in Tables 2-4 under HiddenObject's exact protocol: same optimizer, learning rate, epochs, batch size, 640x480 input, and same train/val splits, then check whether HiddenObject still holds the top mIoU, S_alpha, F_beta, and E_phi values. Separately, drop the thermal or depth channel at inference and measure the performance drop to test whether the claimed modality-agnostic fusion actually preserves detection when a sensor fails.","supporting_citations":[{"cited_title":"Mamba: Linear-time sequence modeling with selective state spaces,","cited_arxiv_id":null,"evidence_quote":"Supplies the selective state-space (Mamba/S6) scan that the encoder, fusion module, and decoder all build on."},{"cited_title":"VMamba: Visual state space model,","cited_arxiv_id":null,"evidence_quote":"Defines Vision Mamba and the SS2D four-direction cross-scan/cross-merge used by the dual-stream encoder and decoder."},{"cited_title":"Sigma: Siamese mamba network for multi-modal semantic segmentation,","cited_arxiv_id":null,"evidence_quote":"Sigma is the closest prior Mamba-based multimodal segmentation baseline on MFNet/PST900 with the VMamba-S backbone, and the fusion module is inspired by it."},{"cited_title":"Cmx: Cross-modal fusion for rgb-x semantic segmentation with transformers,","cited_arxiv_id":null,"evidence_quote":"CMX is the source of the multi-scale cross-modal fusion idea and supplies the AdamW training recipe adopted in the implementation."},{"cited_title":"Mfnet: Towards real-time semantic segmentation for autonomous vehicles with multi-spectral scenes,","cited_arxiv_id":null,"evidence_quote":"MFNet provides one of the RGB-thermal datasets and the early-fusion baseline that the paper must beat."},{"cited_title":"Pst900: Rgb-thermal calibration, dataset and segmentation network,","cited_arxiv_id":null,"evidence_quote":"PST900 provides the second RGB-thermal benchmark and the PSTNet baseline."},{"cited_title":"Delivering arbitrary-modal semantic segmen- tation,","cited_arxiv_id":null,"evidence_quote":"CMNeXt is the strongest RGB-D semantic segmentation baseline on NYU Depth V2 and SUN RGB-D, including the MiT-B4 model that narrowly beats the paper on NYU."}],"review_version":1}