{"id":"de4a598f-c408-4f3f-9462-5cf30def7634","arxiv_id":"2501.00426","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"B2Net, a boundary-aware network that reuses a boundary module and fuses boundary cues across scales, reports state-of-the-art camouflaged object detection results on three public benchmarks.","lead":"B2Net is a neural network for camouflaged object detection that reuses a boundary-aware module at two stages and fuses boundary cues across scales, reporting top scores on COD10K, CAMO, and NC4K benchmarks. A smart generalist might read it to see whether yet another boundary-guided architecture actually moves the needle on a hard segmentation task.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claimed universal superiority is not yet established: with no code, seeds, or significance tests, and margins as small as 0.4 percentage points over FSPNet, the Table I comparison may not reflect protocol-matched runs.","rationale":"I read the paper as making a concrete empirical claim: B2Net achieves the best published numbers on three COD test sets across four metrics. The load-bearing assumption is that the comparison is fair and reproducible. The reader's weakest_assumption identifies exactly this: protocol mismatch and missing artifacts. My stress-test confirms that the margins are small enough that seed-level noise or evaluation-script differences could change the ranking, and the manuscript offers no way to check this because code, weights, and seed settings are absent. I also found a supporting internal wrinkle: the ablation text attributes part of the gain to CBFM using an M2-vs-M5 comparison that is confounded by the addition of the second BAM, and CHAMELEON is claimed as a benchmark but omitted from the results. These are additional reasons to keep the verdict conditional, but they do not amount to a demonstrated flaw in the architecture. The method is internally coherent, the ablations trend in the expected direction, and the reported gains are plausible. Therefore I do not recommend changing the reader's CONDITIONAL verdict; I recommend holding the same conditional posture until the artifacts and protocol details are released.","tokens_in":11475,"tokens_out":4025,"duration_ms":41196,"concrete_test":"Obtain or reimplement the official B2Net code and train it with five different random seeds on the stated 4,040-image training set (CAMO-train + COD10K-train), then evaluate on COD10K-Test and NC4K-Test with the standard evaluation code. Independently run FSPNet's released code under the same input size (352x352), same evaluation script, and, ideally, the same training images. If the B2Net-vs-FSPNet S_alpha gap on NC4K (0.882 vs 0.878) is smaller than the pooled standard deviation across the seeds, the claim of full-metric superiority is not statistically supported and Table I should report variance or use significance tests.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is quantitative: B2Net outperforms 15 published methods on all metrics in Table I. That claim requires that the numbers in Table I come from comparable training and evaluation protocols and that the reported margins exceed run-to-run variability. The paper does not provide code, weights, random seeds, error bars, or statistical significance tests. The concrete margins over the second-best method, FSPNet, are thin: on NC4K-Test, S_alpha is 0.882 vs 0.878 (0.4 pp) and E_phi is 0.945 vs 0.937 (0.8 pp); on Camo-Test, E_phi is 0.935 vs 0.928 (0.7 pp). These differences are within the range of typical seed-to-seed variation for COD training, especially without specifying how edge ground truths are generated for the Dice losses in Eq. 6, which directly affects the BAM supervision. In addition, Section IV.B says experiments were conducted on four datasets, but Table I only reports CAMO, COD10K, and NC4K; CHAMELEON is listed as a benchmark but omitted from results. Finally, the CBFM ablation is partially confounded: the text credits CBFM using the M2-vs-M5 comparison, but M5 differs from M2 by both CBFM and the second BAM. These issues do not prove the method is wrong, but they mean the central claim is not yet independently verifiable from the manuscript alone.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes B2Net, a camouflaged object detection network built on PVTv2 and composed of a Residual Feature Enhanced Module (RFEM), a Boundary Aware Module (BAM) applied twice, and a Cross-scale Boundary Fusion Module (CBFM). The method is trained on 4,040 images from the CAMO and COD10K training sets and evaluated on COD10K-Test, Camo-Test, and NC4K-Test under four metrics (S_alpha, E_phi, F_beta^w, M). The central claim, stated in Section IV.C and Table I, is that B2Net outperforms 15 published methods on all three datasets under all four metrics. An ablation study (Table II) reports monotone gains from adding each module, and Table III shows that the boundary-reuse strategy improves two existing methods. The paper also includes visual comparisons and failure cases.","tokens_in":11769,"tokens_out":3838,"duration_ms":35136,"significance":"If the reported numbers are reproducible, B2Net makes a modest but legitimate contribution to boundary-guided camouflaged object detection. The idea of reusing a boundary-aware module after feature fusion to refine edge semantics is simple and potentially useful, and the ablation results are internally consistent: each added module improves performance across all three datasets, and the strategy transfers to other backbones or methods (Table III). The paper also provides a failure-case analysis, which is a strength. However, the central quantitative claim is not yet independently verifiable from the manuscript alone: no code, weights, random seeds, or statistical significance tests are provided, and the reported margins over the second-best method FSPNet are as small as 0.4 percentage points in S_alpha on NC4K-Test. These caveats do not imply the method is wrong, but they do mean that the current evidence is not sufficient to fully establish the claimed universal superiority.","major_comments":[{"comment":"The paper states in Section IV.B that experiments were conducted on four benchmark datasets (CAMO, CHAMELEON, COD10K, NC4K), but Table I reports results only on COD10K-Test, Camo-Test, and NC4K-Test. CHAMELEON is listed but no results are given, and the abstract claims only three datasets. This inconsistency leaves a stated benchmark unreported and makes the claimed 'three popular COD datasets' unclear. Please either report CHAMELEON results or correct the text to three datasets.","section":"IV.B"},{"comment":"The central claim that B2Net outperforms all 15 methods depends on the comparability of the baseline numbers and on run-to-run variability. The paper does not provide code, weights, random seeds, error bars, or significance tests, and it does not state whether the baseline numbers are copied from the original papers or re-evaluated under the same training and evaluation protocol as B2Net. Since the margins over FSPNet are as small as 0.4 percentage points in S_alpha on NC4K-Test, the ranking could change with seed variation or protocol mismatch. Please supply at least three-seed variance estimates for the main results, and clarify the exact source and protocol of the baseline numbers.","section":"IV.C, Table I"},{"comment":"In the 'Effectiveness of CBFM' paragraph, the comparison between M2 and M5 is confounded because M5 differs from M2 by the addition of both CBFM and a second BAM module. The reported improvement in F_beta^w (5.63% on COD10K) cannot therefore be attributed to CBFM alone. Use M1 versus M3 (CBFM alone) or M4 versus M5 (second BAM alone) to isolate each contribution, and rephrase the corresponding claim.","section":"IV.D"},{"comment":"The loss function in Eq. (6) includes Dice losses on the edge predictions e_i, but the paper does not specify how the edge ground truths G_e are generated during training. Whether they come from dataset-provided edge annotations, from morphological operations on the object masks, or from some other procedure directly affects the BAM supervision and hence the reported boundary quality. Please describe the edge-ground-truth generation process precisely.","section":"III.E, Eq. (6)"}],"minor_comments":[{"comment":"The abstract says 'three challenging benchmark datasets' while Section IV.B says 'four publicly available camouflage object detection benchmark datasets' and lists CHAMELEON. Align the dataset count and the list of used benchmarks.","section":"Abstract and IV.B"},{"comment":"The text says 'we use five evaluation metrics' but then lists only four (S_alpha, E_phi, F_beta^w, M). Either add the missing metric or correct the count to four.","section":"IV.B"},{"comment":"Equation (6) is not typeset clearly in the submitted PDF and appears as an incomplete expression. Ensure the final version has a complete, properly formatted equation with all terms.","section":"III.E"},{"comment":"Some entries have inconsistent decimal precision, e.g., ZoomNet on Camo-Test is listed as S_alpha = 0.82 while other values use three decimals. Use uniform formatting for all table entries.","section":"Table I"},{"comment":"Reference [3] is cited for both PraNet and ZoomNet, but these are different works (PraNet: Fan et al., MICCAI 2020; ZoomNet: Pang et al., CVPR 2022, reference [13] in the text). Please correct the reference list and citations to avoid ambiguity.","section":"References"},{"comment":"The caption mentions 'FAPNet' among the compared methods, but FAPNet does not appear in the method list in Section IV.C or in Table I. Verify the caption and ensure the figure labels match the described comparisons.","section":"Fig. 5 caption"}],"recommendation":"major_revision","confidential_remarks":"This is a fairly standard COD paper whose main contribution is the boundary-reuse idea and the cross-scale fusion of boundary cues. The central claim is plausible and the ablation results are internally consistent, but the manuscript currently lacks the reproducibility evidence (code, seeds, protocol details) and has a confounded ablation comparison in the text. The authors should be asked to address the four major comments before resubmission. If the journal has a strict policy on code sharing, the editor may want to consider whether 'code will be made publicly available' without a URL is sufficient at this stage."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This is a straightforward incremental COD paper: PVTv2 backbone, a Res2Net-style residual enhancement module, a boundary-aware module reused twice, and a top-down cross-scale fusion that injects boundary features. That combination is new, and the ablations are internally consistent—each module adds a monotone gain, and the reuse of BAM after fusion is a genuinely simple idea that seems to help, including when transplanted onto BGNet and BSANet. The paper also includes failure cases and a figure showing the second boundary output is cleaner than the first, which is more honest than most in this subfield. Credit where due: the architecture is clearly described, the ablation design mostly isolates the contributions, and the comparison set of 15 methods is current.\n\nThe soft spots are the usual ones for this area, but they matter more because the headline claim is purely quantitative. Table I shows B2Net beating FSPNet by 0.4–0.8 percentage points on several key metrics. That is within typical seed-to-seed variance for COD training, and the paper gives no error bars, no significance tests, and no code or weights. The text says experiments were run on four datasets, but CHAMELEON is listed and never reported anywhere. The ablation narrative also has a small confound: the CBFM credit uses M2 vs M5, but M5 differs from M2 by both CBFM and the second BAM, so the two are not cleanly separated. There are minor citation/notation errors (FAPNet appears in a figure caption without definition, ZoomNet is mis-cited). None of these sink the paper, but together they mean the central \"outperforms all 15 methods\" claim is not yet something I can verify from the manuscript alone.\n\nWho is this for? Someone working on COD or boundary-guided segmentation who wants one more data point on the reuse-boundary-late trick. It is not a field re-organizer. The right call is to send it to peer review, but with a strong request for code/weights, error bars or at least multiple seeds, CHAMELEON numbers, and a corrected ablation table. If the margins hold under protocol-matched runs, it is a publishable workshop-to-conference-level result. If they don't, the paper still has a useful module and a clean ablation story.","headline":"A solid incremental COD paper with a defensible architecture, but the reported margins over FSPNet are thin and the missing code/seeds make the headline claim not yet independently verifiable.","tokens_in":12302,"tokens_out":808,"would_cite":false,"duration_ms":10147,"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":"B2Net claims that running a boundary-aware module twice, with a cross-scale fusion cascade in between, produces sharper camouflaged-object boundaries and outperforms 15 published methods on three standard benchmarks.","keywords":["camouflaged object detection","boundary awareness","boundary reuse","cross-scale fusion","edge guidance","feature enhancement","PVTv2 backbone","COD benchmark"],"falsifier":"Independently re-run the described training recipe (CAMO plus COD10K training subsets, 352x352 inputs, Adam with initial learning rate 8e-5, 100 epochs) and recompute the four metrics on COD10K-Test, Camo-Test, and NC4K-Test; the central claim fails if the reported $S_\\alpha$ values of 0.862, 0.866, and 0.882 are not reproduced within a small tolerance.","tokens_in":11248,"feed_emoji":"🦎","tokens_out":9199,"duration_ms":82234,"temperature":0.7,"pith_summary":"The paper argues that boundary-guided camouflaged-object detection fails when edge priors are produced only once, early in the network, because those early priors are noisy and mislead the segmentation. To fix this, B2Net runs a boundary aware module twice: once on enhanced multi-scale features to get an initial boundary, and again after a cross-scale boundary fusion cascade has blended boundary cues into object features, producing a refined boundary that is then concatenated with the fused features. The network is completed by a residual feature enhanced module that enriches each backbone level before the boundary modules see it. The paper reports that this two-pass boundary design outperforms 15 published methods on all four standard metrics on COD10K-Test, Camo-Test, and NC4K-Test, with a three-dataset average $S_\\alpha$ gain of 0.93% over the second-best method.","feed_headline":"Reusing boundaries twice sharpens camouflaged object detection","feed_subtitle":"B2Net reports top scores across three COD benchmarks by refining edge cues a second time after cross-scale fusion.","key_machinery":"The load-bearing device is the Boundary Aware Module (BAM), a compact block that adds the two low-level features, multiplies the sum by a high-level semantic feature, and then applies convolutions, a skip connection, max pooling, and spatial attention to keep only object-related edges. The first BAM produces a preliminary edge map; the second BAM, run on the output of the Cross-scale Boundary Fusion Module (CBFM), produces the refined edge map that is concatenated with the fused object features. CBFM carries the fusion: for each shallow level it multiplies the feature by the current edge feature with a learnable weight $\\alpha$, concatenates it with the upsampled output of the previous fusion level, and refines the result, so boundary information is propagated downward across scales. A Residual Feature Enhanced Module (RFEM), inspired by Inception and Res2Net, widens each backbone feature with four residual branches before the boundary modules.","core_discovery":"The central claim is that reusing the same boundary-aware module after cross-scale fusion, rather than generating an edge map once at the start, yields both a more accurate boundary and a better camouflaged-object segmentation. In B2Net, the first BAM takes low-level spatial features and high-level semantic features to predict an initial edge; CBFM modules then inject that edge into the object features in a top-down cascade; and a second BAM, reading the fused output, predicts a refined edge that is concatenated with the cascade's features for the final prediction. The paper reports this design reaches $S_\\alpha = 0.862$, $0.866$, $0.882$ and $M = 0.023$, $0.048$, $0.033$ on COD10K-Test, Camo-Test, and NC4K-Test respectively, and the ablations attribute the gains to the BAM, the CBFM, and especially the second boundary pass.","pith_inferences":["If the Table I protocol is reproducible, the most direct next stress test is tiny-object scenes, which the paper itself identifies as the main failure mode.","The two-pass boundary idea could be tried as a drop-in modification on other COD models that use a single early edge prior; the paper only demonstrates it on BGNet and BSANet.","Because the second BAM sees already-fused features, the benefit should be largest when the first edge prior is wrong; a synthetic benchmark with corrupted or degraded boundary supervision could test this prediction.","The paper's 352x352 input resolution leaves open whether the boundary-reuse gains persist at higher resolutions where fine edge detail is better preserved."],"forward_implications":["The paper's central design principle is that boundary-guided COD should regenerate edge predictions after fusion instead of relying on an early edge prior.","Reported gains over the second-best method (average $S_\\alpha$ +0.93%, $E_\\phi$ +0.68%, $F_\\beta^\\omega$ +2.5%) indicate that a second boundary pass gives measurable improvements on standard benchmarks.","Table III shows the boundary-reuse strategy is transferable: adding a second BAM to BGNet and BSANet improves their scores on the three test sets.","The multi-loss supervision (weighted IoU, weighted BCE, and Dice on edges at three side outputs) trains the whole two-boundary network in 100 epochs, so the added complexity is modest."],"supporting_citations":[{"why":"Supplies the COD10K benchmark used for training and testing and the SINet baseline that defines the search-identification paradigm.","marker":"[6]"},{"why":"Provides the PVTv2 backbone that extracts the four-scale features feeding the RFEM, BAM, and CBFM pipeline.","marker":"[11]"},{"why":"Represents the boundary-guided approach whose early edge-prior limitation motivates the paper's boundary-reuse design.","marker":"[9]"},{"why":"Represents the boundary-guided separated-attention baseline whose single early edge prior is compared against the proposed method.","marker":"[10]"},{"why":"Supplies the CAMO dataset used as part of the 4,040-image training set and as a test benchmark.","marker":"[18]"},{"why":"Supplies the NC4K test benchmark used in the main comparison table.","marker":"[20]"},{"why":"Provides the Res2Net multi-scale block that inspires the residual branch design of RFEM.","marker":"[15]"},{"why":"Supplies the spatial attention mechanism used inside the Boundary Aware Module.","marker":"[16]"}],"fun_headline_variants":["Second boundary pass sharpens camouflaged object detection","B2Net: Double boundary pass yields state-of-the-art COD","Repeated boundary awareness improves COD accuracy","Two-stage boundary refinement wins camouflaged detection"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The comparison with 15 prior methods is only as strong as the shared protocol behind the baseline numbers in Table I, since the paper provides no code, no error bars, and no significance tests to confirm that the reported one-to-two-point margins would replicate.","fun_headline_variants_meta":{"raw":{"variants":["Second boundary pass sharpens camouflaged object detection","B2Net: Double boundary pass yields state-of-the-art COD","Repeated boundary awareness improves COD accuracy","Two-stage boundary refinement wins camouflaged detection"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000454,"raw_usage":{"total_tokens":2293,"prompt_tokens":964,"completion_tokens":1329,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":580,"completion_tokens_details":{"reasoning_tokens":1267}},"tokens_in":580,"tokens_out":1329,"duration_ms":10722,"temperature":1.0,"reasoning_tokens":1267,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T22:50:54.388453+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Independently re-run the described training recipe (CAMO plus COD10K training subsets, 352x352 inputs, Adam with initial learning rate 8e-5, 100 epochs) and recompute the four metrics on COD10K-Test, Camo-Test, and NC4K-Test; the central claim fails if the reported $S_\\alpha$ values of 0.862, 0.866, and 0.882 are not reproduced within a small tolerance.","supporting_citations":[{"cited_title":"Camouflaged object detection","cited_arxiv_id":null,"evidence_quote":"Supplies the COD10K benchmark used for training and testing and the SINet baseline that defines the search-identification paradigm."},{"cited_title":"Pvt v2: Improved baselines with pyramid vision transformer","cited_arxiv_id":null,"evidence_quote":"Provides the PVTv2 backbone that extracts the four-scale features feeding the RFEM, BAM, and CBFM pipeline."},{"cited_title":"I can find you! boundary-guided separated attention network for camouflaged object detection","cited_arxiv_id":null,"evidence_quote":"Represents the boundary-guided separated-attention baseline whose single early edge prior is compared against the proposed method."},{"cited_title":"Anabranch network for camouflaged object segmentation","cited_arxiv_id":null,"evidence_quote":"Supplies the CAMO dataset used as part of the 4,040-image training set and as a test benchmark."},{"cited_title":"Simultaneously localize, segment and rank the camouflaged objects","cited_arxiv_id":null,"evidence_quote":"Supplies the NC4K test benchmark used in the main comparison table."},{"cited_title":"Res2net: A new multi -scale backbone architecture","cited_arxiv_id":null,"evidence_quote":"Provides the Res2Net multi-scale block that inspires the residual branch design of RFEM."},{"cited_title":"Cbam: Convolutional block attention module","cited_arxiv_id":null,"evidence_quote":"Supplies the spatial attention mechanism used inside the Boundary Aware Module."}],"review_version":1}