{"id":"7670afc4-dd87-441c-828e-6761cf4291c6","arxiv_id":"2505.09168","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"DRRNet is a four-stage camouflaged object detection network that fuses global and local features and then applies two rounds of reverse refinement to sharpen object boundaries.","lead":"The paper introduces DRRNet, a neural network for camouflaged object detection that combines global scene context with fine local details and refines results through two reverse passes. It reports top results on three standard camouflage benchmarks and shows the same network can also segment colon polyps.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported SOTA gains in Table I are within run-to-run noise (e.g., CAMO S-alpha 0.881 vs 0.881; no CIs, seeds, or paired tests), and the Section IV.B fairness claim is unverifiable; 'significantly outperforms' is not established.","rationale":"The reader's conditional verdict is appropriate. The weakest point is indeed the empirical comparison, but I would sharpen it: the fairness premise is only part of the problem. Even if every baseline were reproduced exactly as in its original paper, the differences in Table I are so close to zero that the word 'significantly' requires statistical support that the paper does not provide. This is a standard correctness risk, not an accusation: the reported numbers may be honest, but they cannot carry the stated conclusion until paired error intervals are supplied. The ablation studies are systematic and the architecture is described in sufficient detail to be reimplemented, which supports a conditional rather than reject verdict. The promised code without a commit hash or released weights currently prevents independent verification. I did not find an internal inconsistency that would invalidate the method itself; the concern is entirely about the size and verifiability of the claimed empirical margin.","tokens_in":19659,"tokens_out":7243,"duration_ms":67781,"concrete_test":"Run DRRNet and the closest baselines (EFNet, BDCL-Net, PRBENet) on CAMO, COD10K, and NC4K using each method's official code and checkpoints; train DRRNet with at least five seeds and compute paired per-image deltas for S-alpha, E-phi, F-beta-w, and MAE with bootstrap 95% confidence intervals. Also verify that reproduced baseline numbers match the values in their original papers. If any bootstrap interval contains zero, or any reproduced baseline exceeds the Table I entry by more than 0.001, the headline 'significantly outperforms' should be softened to 'comparable' and Table I recomputed.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is empirical: DRRNet 'significantly outperforms all SOTA methods across all datasets.' The decisive evidence in Table I consists of point estimates whose largest win over the closest competitor is often 0.001-0.004 in S-alpha/E-phi, with exact ties on CAMO S-alpha and NC4K S-alpha (0.881 vs 0.881 and 0.896 vs 0.896). No variance, confidence interval, paired significance test, or multi-seed result is reported. Retraining a COD network with a different seed typically shifts S-alpha by more than these margins, so the reported deltas are consistent with the null hypothesis that DRRNet is not better than EFNet or BDCL-Net. The fairness sentence in Section IV.B ('all prediction maps are generated using identical code, and all metrics are computed with the same implementation') does not resolve this: it does not state whether each baseline used its official released checkpoint and inference code, or whether predictions were regenerated under a unified pipeline. If any baseline was disadvantaged by non-official weights or modified inference, its Table I numbers could be depressed by more than the claimed advantage. Thus the empirical superiority clause of the central claim is unsupported even though the architectural components may be sound.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes DRRNet, a four-stage camouflaged object detection (COD) architecture built around an Omni-Context Module (OCM) for global context, a Micro Detail Module (MDM) for local details, a Macro-Micro Fusion (MMF) module for spatial-frequency fusion, a Global Rough Decoder (GRD), and a Dual Reverse Refinement Module (DRRM) that performs two rounds of spatial-edge and frequency-domain refinement. The method is evaluated on CAMO, COD10K, and NC4K against 17 published methods, with additional experiments on four polyp segmentation datasets. The authors report that DRRNet outperforms all compared state-of-the-art methods on nearly all metrics, and they provide ablation studies for each architectural component. The central claim is that DRRNet 'significantly outperforms all SOTA methods across all datasets' without post-processing.","tokens_in":19930,"tokens_out":4955,"duration_ms":49563,"significance":"If the reported numbers are reproducible and the comparisons are fair, DRRNet would be a competitive COD architecture with a clear and well-motivated design. The paper follows standard COD evaluation practice, uses three widely adopted benchmarks and four metrics, ablates every proposed module, reports parameter and FLOP counts, and releases code. The modular decomposition into global-context, local-detail, fusion, and reverse-refinement components is sensible, and the polyp segmentation transfer is a useful generalization check. The principal weakness is that the headline claim of significant state-of-the-art improvement rests on point-estimate differences of 0.001–0.004 in several metrics, with exact ties on two key entries, and no variance, confidence intervals, or paired significance tests are reported. As a result, the empirical superiority clause is not currently established, even though the architecture itself appears sound and the ablations support the qualitative contribution of each module.","major_comments":[{"comment":"The claim that 'DRRNet significantly outperforms all SOTA methods across all datasets' is not supported by the reported evidence. In Table I, the largest margins over the closest competitor are typically 0.001–0.004 in Sα and Eϕ, and there are exact ties: CAMO Sα is 0.881 for both DRRNet and EFNet, and NC4K Sα is 0.896 for both. No confidence intervals, standard deviations across training seeds, or paired statistical tests are provided. Since COD training is stochastic and run-to-run variation in Sα is typically larger than these margins, the reported numbers are consistent with the null hypothesis that DRRNet is not better than EFNet or BDCL-Net. The authors should add multi-seed results with variance reporting or paired significance tests across the test sets, or alternatively soften the claim to describe the observed point estimates.","section":"Section IV.B, Table I"},{"comment":"The fairness statement 'all prediction maps are generated using identical code, and all metrics are computed with the same implementation' is ambiguous and does not establish that the 17 baselines were not disadvantaged. It is not stated whether each baseline used its official released checkpoint and official inference code, or whether all baseline predictions were regenerated under a unified pipeline that may not match each model's original inference procedure. If any baseline was re-implemented, re-trained, or run with non-official weights, its Table I numbers could be depressed by more than the claimed advantage. The authors should specify, for every baseline, the source of the predictions (official checkpoint and code vs. re-implementation), and report the exact protocol used. This is load-bearing because the central claim of SOTA improvement depends on the fairness of these comparisons.","section":"Section IV.B"},{"comment":"The polyp segmentation extension is presented as evidence of generalization, but the claim that DRRNet 'significantly outperforms existing methods' is again based on single-run point estimates with no variance or significance testing. For example, on CVC-300 the reported MAE improvement over GLCONet is 0.009 to 0.007, and on Kvasir the Sm gap is 0.925 vs. 0.872 for GLCONet; these are large in some cases, but no error bars or statistical tests are provided. This is a secondary claim relative to the COD result, but it should be reported with the same caveats or supported by appropriate statistics.","section":"Section IV.D, Table VII"}],"minor_comments":[{"comment":"Equation (5) has a mismatched parenthesis: the expression 'li = CBR(Φ(SE(cat(Faspp,Fdw),Fres))+x0' appears to be missing a closing parenthesis after the CBR argument. Please fix the equation for clarity.","section":"Section III.B, Eq. (5)"},{"comment":"In Eq. (6), the spatial branch is written as 'xspatial = CBR(Φ(gi,li))', but the textual description says the GroupFusionBlock operates on each channel group separately. This notational mismatch makes the actual grouping operation ambiguous. Also, the expression for xfreq is missing a closing parenthesis after the modulation term.","section":"Section III.C, Eq. (6)"},{"comment":"The text cites [32] as the Pyramid Vision Transformer (PVTv2) backbone, but the reference listed is an Applied Sciences paper titled 'Camouflaged Object Detection That Does Not Require Additional Priors.' This appears to be the wrong reference; the correct PVTv2 citation should be provided.","section":"References, [32]"},{"comment":"The sentence '...which significantly improving the accuracy of camouflaged object segmentation' contains a grammatical error; it should read '...which significantly improves the accuracy...'","section":"Section III.A"},{"comment":"The checkmark layout of Table II is difficult to read: the header includes 'Baseline' alongside the five module columns, and it is not immediately clear which row corresponds to the baseline and which row corresponds to the full model. Please use explicit row labels such as 'Baseline', '+OCM', '+MDM', etc., and state in the caption what each checkmark denotes.","section":"Table II"}],"recommendation":"major_revision","confidential_remarks":"The paper's architectural contribution is plausible and the ablations are informative, but the headline empirical claim needs statistical support and a verifiable comparison protocol before the paper can be accepted. In addition, the incorrect PVTv2 citation in [32] should be corrected; it suggests a need for a careful reference check. The paper fits the scope of a computer-vision journal and, with the requested revisions, could become a solid contribution."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The architecture is real and the experiments follow field norms, but the central claim that DRRNet “significantly outperforms all SOTA methods” is not established by Table I. The largest wins over the strongest baselines are often 0.001–0.004 in S-alpha/E-phi, with exact ties on CAMO and NC4K S-alpha (0.881 vs 0.881, 0.896 vs 0.896). There are no error bars, no multiple seeds, no paired significance tests. In COD, rerunning a strong baseline with a different seed typically moves S-alpha by more than these margins. The Section IV.B fairness sentence says all predictions were generated with identical code and metrics with the same implementation, but it does not state whether each baseline used its official released checkpoint and inference code. So the empirical superiority clause is unverified, even though the components may be sound.\n\nWhat is genuinely new: the DRRNet pipeline—OCM global context, MDM local detail, MMF gated spatial/frequency fusion, GRD coarse decoding, and DRRM dual reverse refinement—is a specific combination I do not recall seeing before. The ablations in Table II show each module helps, the polyp segmentation transfer in Table VII is a useful generalization test, and the params/FLOPs comparison is informative. The paper is clearly written overall and the limitations paragraph is honest.\n\nThe soft spots are real but addressable. Besides the missing statistics, the references are sloppy: [13] is an eye-tracking paper, not a COD method; [32] is cited for PVTv2 but listed as a camouflaged detection paper; [55] duplicates [30]. Equation (6) uses gamma_GFB without defining it, and the DRRM text confuses Fc/Fs/Ff in the frequency branch. No code hash or weights are provided, so exact reproducibility cannot be checked.\n\nThis is a useful paper for COD specialists, and the architecture is likely to become a citeable baseline even if the SOTA claim does not survive scrutiny. I would send it to peer review rather than desk-reject, with a request for multi-seed variance or confidence intervals, a clarified baseline-reproduction protocol, and cleanup of the citations and notation. Serious referee effort would improve it; the core contribution is legitimate.","headline":"A coherent COD architecture paper whose headline SOTA claim is not backed by the reported numbers; worth referee time as a methods contribution, but the empirical superiority statement needs to be softened and re-supported.","tokens_in":20440,"tokens_out":1895,"would_cite":false,"duration_ms":21753,"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":"DRRNet pairs global and local feature fusion with two-stage reverse refinement to claim top scores on CAMO, COD10K, and NC4K.","keywords":["camouflaged object detection","feature fusion","frequency-domain calibration","reverse refinement","edge-aware segmentation","pyramid vision transformer","spatial-frequency modeling","polyp segmentation"],"falsifier":"An independent rerun of CAMO, COD10K, and NC4K using each competitor's official released prediction maps or official inference code, with the metric implementation fixed, would settle the claim: if EFNet, MVGNet, or BDCL-Net then matches or exceeds DRRNet on S-alpha or weighted F-measure, the reported state-of-the-art margin disappears. The paper's own conclusion names dynamic backgrounds and overlapping multi-target instances as remaining weak spots, so a stress benchmark built from video frames or densely overlapping camouflaged objects offers a second disconfirmation test.","tokens_in":19482,"feed_emoji":"🦎","tokens_out":7100,"duration_ms":66865,"temperature":0.7,"pith_summary":"This paper introduces DRRNet, a four-stage network for camouflaged object detection built around a context-detail-fusion-refinement pipeline. It claims that separating global scene understanding (macro) from local microstructure cues (micro), fusing them in both spatial and frequency domains, and then applying two successive rounds of reverse refinement to a coarse prediction produces masks with stronger boundary continuity and less background interference. On the three standard benchmarks (CAMO, COD10K, NC4K), DRRNet reports the best or tied-best score on nearly every metric among 17 compared methods, without post-processing. A transfer experiment to polyp segmentation suggests the design generalizes to other low-contrast segmentation tasks. The practical payoff would be a single architecture that handles both the 'lost hair-like detail' failure of global-only models and the 'vegetation-pattern distraction' failure of local-only models.","feed_headline":"DRRNet beats all 17 baselines on three camouflage benchmarks","feed_subtitle":"Global context plus local detail, then two rounds of inverse refinement, lift accuracy on CAMO, COD10K and NC4K.","key_machinery":"The load-bearing mechanism is the Dual Reverse Refinement Module (DRRM), applied twice in the decoder. Given coarse predictions from deeper stages $O_{i+1}$ and $O_{i+2}$, it forms inverted residual maps $R_1 = 1-\\sigma(O_{i+1})$ and $R_2 = 1-\\sigma(O_{i+2})$, uses their sum to weight the fused feature map, and adds the coarse outputs back as residual connections. In parallel, a frequency branch takes the Fourier transform of the fused features, modulates the spectrum with convolution-derived weights, and transforms back to the spatial domain. This dual spatial-frequency calibration, repeated in two stages, is what the paper credits for suppressing background noise while preserving boundary continuity.","core_discovery":"The central claim is that the longstanding COD trade-off between global context and local detail can be resolved architecturally rather than by choosing one side: DRRNet obtains panoramic context through a multi-branch Omni-Context Module, recovers microstructural cues through an ASPP-style and depthwise-separable Micro-Detail Module, and merges the two streams in a Macro-Micro Fusion module that jointly models spatial and frequency representations. The decoder first produces a coarse global prediction, then a Dual Reverse Refinement Module applies two inverse-refinement rounds, each using spatial edge priors from inverted sigmoid maps and frequency-domain noise suppression, to progressively remove background false positives and sharpen boundaries. Supported by ablations in which each module is removed or replaced, the paper argues that the components contribute additively and that the two-stage reverse refinement is what converts a rough localization into a precise segmentation. The paper's stated conclusion is that DRRNet outperforms all 17 compared state-of-the-art methods on CAMO, COD10K, and NC4K across S-measure, E-measure, weighted F-measure, and MAE without any post-processing.","pith_inferences":["The paper does not run a direct comparison against the officially released prediction maps of its closest competitors; a natural test is to re-evaluate on those maps to see whether the roughly 0.001-point S-alpha gaps persist or shrink to noise.","The paper's own closing caveat—dynamic backgrounds and overlapping multi-target instances remain hard—suggests the next stress test should be a video or multi-instance camouflage benchmark, where the current single-image metrics may not reflect the claimed robustness.","Because the ablation tables show metric gains of only a few thousandths on some settings, an independent rerun with multiple random seeds would clarify which modules carry the real signal; the paper reports single-run comparisons.","The frequency-domain mechanism works by modulating Fourier amplitude coefficients, so a plausible inference is that high-frequency texture noise in vegetation-like backgrounds is the main source of false positives DRRNet suppresses; this could be tested by adding synthetic high-frequency noise to test images and measuring MAE degradation."],"forward_implications":["If the reported numbers hold, DRRNet becomes the new reference point on CAMO, COD10K, and NC4K, with the highest weighted F-measure on COD10K (0.817) and the lowest MAE (0.019) among the compared methods.","The two-stage reverse refinement recipe—coarse decode, then invert and re-weight with edge priors—can be lifted out and attached to other encoder-decoder detectors as a plug-in decoder.","The spatial-plus-frequency fusion pattern gives later COD designs a concrete alternative to stacking more Transformer blocks: the gains here come from explicit dual-domain calibration rather than larger receptive fields.","Because the same network transfers to polyp segmentation with only retraining, low-contrast medical segmentation tasks are a direct application area if the COD results replicate."],"supporting_citations":[{"why":"Supplies the COD10K dataset, the training/test split, and the foundational SINet baseline the comparison builds on.","marker":"[11]"},{"why":"Supplies the CAMO benchmark dataset used for both training and testing.","marker":"[23]"},{"why":"Supplies the NC4K benchmark dataset used for evaluation.","marker":"[24]"},{"why":"Provides the Pyramid Vision Transformer backbone whose multi-scale features feed the macro and micro branches.","marker":"[32]"},{"why":"Inspires the dual-path local detail extraction design and serves as a comparison baseline in the polyp segmentation transfer.","marker":"[33]"},{"why":"Provides the SENet attention mechanism used for channel recalibration inside OCM, MDM, GRD, and DRRM.","marker":"[35]"},{"why":"The closest prior competitor on the benchmarks, providing the best previous scores that DRRNet must beat.","marker":"[21]"},{"why":"Defines the S-measure used to evaluate structural similarity of the predicted foreground maps.","marker":"[48]"}],"fun_headline_variants":["DRRNet's dual reverse refinement beats 17 baselines on 3 COD sets","Macro-micro fusion and dual refinement lift DRRNet past 17 models","DRRNet: context-detail fusion plus two-step inverse refinement wins","Two rounds of reverse refinement help DRRNet top COD benchmarks","DRRNet merges scales and refines twice to beat all 17 methods"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The strongest empirical claim rests on the fairness of the comparisons: the paper states that all prediction maps were generated with identical code, and if the reproduced baselines were not optimally tuned or their official inference code was not used, the few-thousandths-of-a-point gains in Table I are not meaningful.","fun_headline_variants_meta":{"raw":{"variants":["DRRNet's dual reverse refinement beats 17 baselines on 3 COD sets","Macro-micro fusion and dual refinement lift DRRNet past 17 models","DRRNet: context-detail fusion plus two-step inverse refinement wins","Two rounds of reverse refinement help DRRNet top COD benchmarks","DRRNet merges scales and refines twice to beat all 17 methods"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00047,"raw_usage":{"total_tokens":2375,"prompt_tokens":1019,"completion_tokens":1356,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":635,"completion_tokens_details":{"reasoning_tokens":1256}},"tokens_in":635,"tokens_out":1356,"duration_ms":12145,"temperature":1.0,"reasoning_tokens":1256,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T21:37:23.909062+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"An independent rerun of CAMO, COD10K, and NC4K using each competitor's official released prediction maps or official inference code, with the metric implementation fixed, would settle the claim: if EFNet, MVGNet, or BDCL-Net then matches or exceeds DRRNet on S-alpha or weighted F-measure, the reported state-of-the-art margin disappears. The paper's own conclusion names dynamic backgrounds and overlapping multi-target instances as remaining weak spots, so a stress benchmark built from video frames or densely overlapping camouflaged objects offers a second disconfirmation test.","supporting_citations":[{"cited_title":"Fan, G.-P","cited_arxiv_id":null,"evidence_quote":"Supplies the COD10K dataset, the training/test split, and the foundational SINet baseline the comparison builds on."},{"cited_title":"Cam- ouflaged Object Detection That Does Not Require Additional Pri- ors,","cited_arxiv_id":null,"evidence_quote":"Provides the Pyramid Vision Transformer backbone whose multi-scale features feed the macro and micro branches."},{"cited_title":"GLCONet: Learning Multi-source Perception Representation for Camouflaged Object Detection","cited_arxiv_id":"2409.09588","evidence_quote":"Inspires the dual-path local detail extraction design and serves as a comparison baseline in the polyp segmentation transfer."}],"review_version":1}