{"id":"11f6ac12-6edb-41c6-abe7-c774b231740e","arxiv_id":"2608.11135","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"The authors propose a 16,245-image benchmark that includes negative samples for camouflaged object detection and a network that jointly predicts object presence, camouflage presence, and segmentation masks.","lead":"This paper introduces OPC16K, a benchmark for camouflaged object detection that adds pure-background and non-camouflaged images to the usual positive samples, and OPCNet, a network that decides whether a camouflaged object exists before producing a segmentation mask. A generalist might read it to see how the field is trying to make camouflage detectors safe to deploy in unconstrained settings where most images contain no camouflaged object.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table 3 contradicts the paper's central claim: OPCNet's FPRno (0.181) is nearly double VSCode's (0.093), so the stated 'lowest FPRbg and FPRno' is false for NOCOD.","rationale":"The single most load-bearing assumption behind the paper's central claim is that OPCNet really delivers the best negative-sample rejection. The paper states this in the abstract ('significantly reducing false positives on negative samples') and in Main Results ('lowest FPRbg and FPRno'). But Table 3, the paper's own quantitative comparison, contradicts it: OPCNet's FPRno = 0.181 is roughly twice VSCode's 0.093. This is not a subtle calibration issue or a matter of external SOTA; it is an internal inconsistency between prose and table. The reader's weakest_assumption focused on the fairness of adapting baselines to three classes. That concern is legitimate and cannot be checked because the adaptation is deferred to an unpublished appendix. However, the internal contradiction is even more decisive: even if the adaptation were perfectly fair, the paper's claim as written would still be wrong for FPRno. The reported Accno supports the same conclusion (0.817 vs 0.907). The paper could perhaps salvage the contribution by repositioning OPCNet as best overall accuracy and best FPRbg, while acknowledging VSCode is better on NOCOD rejection; but the current text overclaims. Given the absence of released data and code and the unverifiable appendix, the honest verdict remains REJECT.","tokens_in":12452,"tokens_out":5023,"duration_ms":42834,"concrete_test":"Rerun VSCode and OPCNet on the OPC16K NOCOD test subset with the paper's official code and split, convert pixel masks to scene labels by the stated rule, and recompute FPRno. Then audit the appendix's three-class adaptation: confirm VSCode's head was extended to predict a NOCOD class with the same supervision as OPCNet. If the recomputed FPRno remains 0.093 vs 0.181, the main-text claim is false; if the adaptation gave VSCode a privileged NOCOD output, the comparison must be redone symmetrically.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's headline empirical claim ('Main Results', Table 3) is that 'OPCNet achieves ... the lowest FPRbg and FPRno'. Under the paper's own evaluation protocol, FPRno is the false-positive rate on non-camouflaged-object images, i.e., the central negative-sample rejection metric. The table lists VSCode with FPRno = 0.093 and OPCNet with FPRno = 0.181, so OPCNet is not the best on NOCOD rejection; it is second-best, behind VSCode by 0.088. OPCNet also has lower Accno (0.817 vs 0.907). Averaging the two negative FPRs, VSCode is better (0.084 vs 0.116). This is an internal inconsistency: no external baseline, backbone, or adaptation assumption is needed to see that the stated 'lowest FPRbg and FPRno' is false on FPRno. Since false-positive suppression on negative samples is the paper's stated motivation and its claimed advantage, this directly undermines the central claim. The fairness of baseline adaptation (deferred to an appendix) is a further unverifiable premise, but the contradiction in Table 3 is more load-bearing.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper addresses camouflaged object detection (COD) under a realistic open-world protocol, where inputs may contain a camouflaged object (CO), a pure background (BG), or a non-camouflaged object (NOCOD). It introduces OPC16K, a benchmark with over 16,000 images organized into these three classes with deliberate foreground-category and scene-level distribution alignment. It also proposes OPCNet, which adds hierarchical existence reasoning, similarity-aware camouflage relation modeling, and existence-aware feature refinement to a segmentation backbone, and it defines an evaluation protocol with three-way classification accuracy, negative-sample false positive rates, and segmentation metrics. The central empirical claim is that OPCNet achieves the best overall performance and, in particular, the lowest false positive rates on both BG and NOCOD inputs.","tokens_in":12700,"tokens_out":5413,"duration_ms":50377,"significance":"If the claims were fully supported, the paper would make a useful contribution: it identifies a real limitation of closed-world COD evaluation, proposes a benchmark that includes negative samples with distribution controls, and introduces an explicit existence-reasoning mechanism with an ablation study showing that each component contributes. The three-way evaluation protocol, although simple, is a reasonable step toward realistic deployment assessment. However, the headline empirical claim is internally contradicted by the paper's own Table 3, and the comparative evaluation rests on a baseline adaptation procedure that is deferred to an unpublished appendix. The paper also promises but does not currently provide the dataset or code, which limits independent verification. The benchmark and method ideas are valuable, but the presented evidence does not support the stated conclusion that OPCNet significantly reduces false positives on negative samples compared with existing methods.","major_comments":[{"comment":"The text states that 'OPCNet achieves the best overall performance under the proposed open-world COD protocol. It obtains the highest three-way classification accuracy and the lowest FPRbg and FPRno', and the Abstract claims 'significantly reducing false positives on negative samples'. Table 3 contradicts this on the NOCOD axis: OPCNet's FPRno is 0.181, whereas VSCode attains 0.093, so OPCNet is not the best on non-camouflaged-object rejection; its Accno (0.817) is also below VSCode's (0.907). Since false-positive suppression on negative samples is the paper's stated motivation and central advantage claim, this internal inconsistency directly undermines the headline result. The ablation in Table 4 confirms that even the full OPCNet (FPRno 0.181) does not reach VSCode's FPRno.","section":"Main Results (Table 3)"},{"comment":"Every comparative claim in Table 3 depends on the adaptation of binary COD baselines to the proposed three-class setting, but the only description is that the baselines are adapted by 'extending their prediction space to three classes: CO, NOCOD, and BG', with details deferred to an appendix that is not part of the submitted manuscript. It is therefore impossible to verify whether the baselines received a comparable classification head, how image-level labels were derived from pixel predictions, whether thresholds were tuned for each baseline, or whether the adaptation is unbiased. If the adaptation handicaps the baselines, the reported superiority of OPCNet would be an artifact of the adaptation rather than of the method itself; this is a load-bearing reproducibility gap for the central comparison.","section":"Compared Methods (Experimental Setup)"},{"comment":"The overall segmentation metrics (oMAE, oSm, oEm, oFwβ) assign worst scores to misclassified samples, e.g., MAE = 1 and Sm = 0 for samples whose predicted scene label is wrong. This makes these 'overall COD segmentation quality' metrics a direct function of three-way classification accuracy, which is precisely the capability that OPCNet's hierarchical existence reasoning is explicitly designed to improve. The co-prefixed metrics provide a useful check, but the paper's headline open-world metrics therefore conflate classification performance with segmentation quality, and the reported margins on oMAE and oSm should be interpreted with this dependency in mind.","section":"Evaluation Protocol"}],"minor_comments":[{"comment":"The heading 'The Proposed OPC16K Dataest' contains a typo; it should read 'Dataset'.","section":"Dataset Construction"},{"comment":"In the provided text, some consecutive numeric entries in the OPCNet row appear without separating spaces (e.g., '0.8180.8750.701'); the camera-ready table should ensure clear column separation for all cells.","section":"Table 3"},{"comment":"The conclusion repeats the claim that OPCNet reduces false positives on negative samples; this statement should be revised to match the actual FPRno numbers in Table 3 or be accompanied by a clear explanation of why the FPRno comparison with VSCode is not decisive.","section":"Conclusion"},{"comment":"The paper states that code and dataset 'will be released', but they are not currently available; independent verification of the benchmark statistics and of the baseline adaptation procedure is therefore not possible from the submission.","section":"Experimental Setup"}],"recommendation":"reject","confidential_remarks":"The internal contradiction between the main-text claim and Table 3 is, in my view, disqualifying for the current version: the paper's core selling point is negative-sample rejection, and its own numbers show that VSCode has a much lower FPRno. The missing baseline-adaptation details would be a serious issue even if the contradiction were fixed, because all comparative conclusions hinge on that adaptation. I would not invite a revision unless the authors can provide the full adaptation protocol, rerun the comparison, and correct the claims accordingly."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Hi [name],\n\nRead the COD paper. The benchmark is the real contribution: OPC16K gives the subfield a much-needed negative-sample axis, and they've put thought into matching foreground categories and scenes between positives and negatives. That alone is worth something, and the existence-reasoning architecture is a reasonable way to structure the task.\n\nThe problem is the main claim. The paper says OPCNet attains the lowest FPRbg and FPRno. Table 3 shows FPRbg is indeed lowest (0.050 vs 0.074), but FPRno is 0.181, well behind VSCode's 0.093. The average negative FPR also favors VSCode. So the stated advantage on negative-sample rejection—the paper's stated motivation—is not supported by its own numbers. That's not an external artifact; it's internal.\n\nI also have smaller concerns. The baseline adaptation to three classes is relegated to the appendix, so we can't tell whether the comparisons are fair. The protocol assigning worst scores to misclassified samples rewards the classification ability OPCNet is explicitly designed for, which is fine as a protocol but should be presented as such. And data/code are only promised, so the benchmark isn't independently checkable yet.\n\nThe dataset and task framing are still relevant to the COD community, and the method idea is coherent. But the headline empirical claim needs correcting, or the paper's case largely falls apart. I'd send it to referees, mostly to get the dataset scrutinized and the comparison rerun; it shouldn't be accepted as is.\n\nRecommendation: engage, but with a clear request to fix the numbers and release the data.","headline":"The benchmark is a genuine step for COD evaluation, but the paper's headline claim about false-positive suppression is contradicted by its own Table 3, so engagement should come with a demand for corrected numbers.","tokens_in":13253,"tokens_out":2106,"would_cite":false,"duration_ms":19648,"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":"The paper argues that camouflaged-object detection must move beyond the closed-world assumption that every image contains a target, and presents a benchmark and network that explicitly decide whether a camouflaged object exists before…","keywords":["camouflaged object detection","open-world evaluation","negative sample rejection","false positive suppression","benchmark dataset","presence-aware segmentation","hierarchical reasoning"],"falsifier":"Retrain a strong baseline such as RUN or USCNet with the same hierarchical existence-reasoning head used by OPCNet (or an equally capable two-stage head) on OPC16K and check whether OPCNet's margins in three-way accuracy, FPRbg, and FPRno persist; alternatively, test OPCNet on a new set of negative images deliberately drawn from categories and scenes outside OPC16K's aligned distribution and measure whether its false-positive advantage collapses.","tokens_in":12217,"feed_emoji":"🦎","tokens_out":3968,"duration_ms":35673,"temperature":0.7,"pith_summary":"Camouflaged-object detection (COD) models are usually trained and tested under a closed-world assumption: every input image contains a camouflaged object. This paper argues that real deployment does not work that way, because images may be pure backgrounds or contain ordinary, non-camouflaged objects, and it shows that current state-of-the-art models fire false positives on such negatives. To fix this, the paper builds OPC16K, a benchmark of 16,245 images organized into camouflaged-object, pure-background, and non-camouflaged-object subsets with deliberately matched category and scene distributions. It then proposes OPCNet, a network that first localizes objects and then reasons about whether camouflage exists, achieving high three-way classification accuracy and low false-positive rates while preserving segmentation quality on genuine camouflaged objects.","feed_headline":"Camouflaged-object AI learns to say \"nothing here\"","feed_subtitle":"New OPC16K benchmark and OPCNet network let detectors reject backgrounds and ordinary objects while still finding real camouflage.","key_machinery":"The load-bearing mechanism is a hierarchical two-stage decision. Stage one computes an object-presence probability $p_{obj}$ from global pooled features, separating pure backgrounds from images that contain any object. Stage two computes a camouflage-presence probability $p_{pres}$ using the Similarity-Aware Camouflage Relation Module (SACRM), which forms foreground and background prototypes from a rough localization mask, measures per-pixel similarity to the background prototype, and feeds foreground, background, similarity statistics, and foreground-background discrepancy into an MLP. The final label comes from thresholds on both probabilities: background if $p_{obj} \\le \\tau_{obj}$, camouflaged if both probabilities are high, and non-camouflaged otherwise. An existence-aware feature refinement module then multiplies a spatial gate by $p_{obj} \\cdot p_{pres}$ to suppress segmentation responses on negative images.","core_discovery":"On its own OPC16K protocol, OPCNet achieves 90.5% three-way classification accuracy, a 5.0% false-positive rate on pure backgrounds, and an 18.1% false-positive rate on non-camouflaged-object images, outperforming five baselines that are retrained on the same benchmark. The central claim is that CAMOUFLAGED-object detection should be reformulated from pure segmentation into a joint problem of object localization and camouflage existence reasoning, so that the model can explicitly answer 'is there really a camouflaged object here?' before producing a mask.","pith_inferences":["The same two-stage 'is it present, then is it the target kind' recipe could transfer to other detection tasks where negatives dominate, such as rare-species monitoring or industrial defect inspection, by replacing the camouflage-existence stage with a task-specific discrimination cue.","The thresholded decision rule implies an operating point that could be tuned per application; the reported 0.5 thresholds are a default, so a validation-set threshold search might widen or narrow the margin over baselines.","Because the paper converts pixel masks of baselines into image-level labels by largest foreground area, a baseline whose segmentation is spatially scattered could be disadvantaged; a fairer comparison might use a dedicated image-level head trained directly on the three-class labels.","If OPCNet is evaluated on negative images whose categories and scenes are deliberately shifted away from OPC16K's aligned distribution, the false-positive advantage may shrink, revealing how much of the gain comes from distribution matching rather than from camouflage reasoning itself."],"forward_implications":["Existing COD models, when retrained on OPC16K, can reduce false positives on negative samples while roughly maintaining segmentation quality on true camouflaged objects.","The three-way classification protocol (CO, BG, NOCOD) provides a standard way to evaluate whether a COD method can say 'no camouflaged object here' rather than always segmenting something.","The two-stage reasoning structure can be attached to any segmentation backbone, since it only adds a global classifier and a similarity-based relation module on top of existing features.","The dataset's distribution alignment between positive and negative samples discourages shortcut learning through category or scene domain differences, making the benchmark a stricter test of camouflage-specific reasoning."],"supporting_citations":[{"why":"Supplies the COD10K camouflaged-object images and masks that form a large part of OPC16K's positive samples and the benchmark lineage the paper extends.","marker":"(Fan et al. 2020a)"},{"why":"Provides the CAMO dataset, another source of camouflaged-object positives used in OPC16K.","marker":"(Le et al. 2019)"},{"why":"Provides the NC4K dataset used as a camouflaged-object positive source in OPC16K.","marker":"(Lv et al. 2021)"},{"why":"USCNet and USC12K are the closest related joint SOD/COD method and dataset; the paper positions OPC16K as improving on their negative-sample coverage and evaluates against USCNet.","marker":"(Zhou et al. 2025)"},{"why":"RUN is a state-of-the-art COD method whose 97.4% false-positive rate on OPC16K motivates the negative-sample problem and serves as a baseline.","marker":"(He et al. 2025)"},{"why":"CamoDiffusion is a diffusion-based COD baseline evaluated against OPCNet on the proposed protocol.","marker":"(Sun et al. 2025)"},{"why":"SAM3 is used to initially annotate non-camouflaged object masks in OPC16K before manual refinement.","marker":"(Carion et al. 2025)"},{"why":"PVTv2-B4 is the backbone of OPCNet and also of the RUN baseline, making the comparison backbone-consistent.","marker":"(Wang et al. 2022)"}],"fun_headline_variants":["AI that spots camouflage now knows when to say 'none'","Realistic camouflage detection: learning to reject empty scenes","OPCNet: teaching camouflage detectors to admit nothing's there","New benchmark forces camouflage AI to face real-world negatives","From pure segmentation to existence reasoning in camouflage detection"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The comparison of negative-sample rejection rests on the fairness of adapting each baseline to the three-class setting: the paper says baseline prediction spaces were extended to CO, NOCOD, and BG with details deferred to the appendix, so if that adaptation handicaps any baseline, the reported superiority could be an artifact of the adaptation rather than of OPCNet itself.","fun_headline_variants_meta":{"raw":{"variants":["AI that spots camouflage now knows when to say 'none'","Realistic camouflage detection: learning to reject empty scenes","OPCNet: teaching camouflage detectors to admit nothing's there","New benchmark forces camouflage AI to face real-world negatives","From pure segmentation to existence reasoning in camouflage detection"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000205,"raw_usage":{"total_tokens":1392,"prompt_tokens":945,"completion_tokens":447,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":561,"completion_tokens_details":{"reasoning_tokens":368}},"tokens_in":561,"tokens_out":447,"duration_ms":4601,"temperature":1.0,"reasoning_tokens":368,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T05:30:14.872158+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain a strong baseline such as RUN or USCNet with the same hierarchical existence-reasoning head used by OPCNet (or an equally capable two-stage head) on OPC16K and check whether OPCNet's margins in three-way accuracy, FPRbg, and FPRno persist; alternatively, test OPCNet on a new set of negative images deliberately drawn from categories and scenes outside OPC16K's aligned distribution and measure whether its false-positive advantage collapses.","supporting_citations":[],"review_version":1}