{"id":"49682135-880e-49c4-b947-3c882bbe0828","arxiv_id":"2412.18977","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A new COD benchmark with class labels and a text-guided framework that outperforms existing camouflaged object detectors.","lead":"This paper introduces class-guided camouflaged object detection, adding text labels of object classes to the standard image-only task, and releases a new dataset CamoClass with class annotations for about 10,500 images. It also presents CGNet, a two-stage framework that uses CLIP text and image features to improve segmentation, reporting gains over existing methods.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Dataset totals in §3 and §5.1 are inconsistent, and the claimed 'average 5.5% improvement' in §5.2 is not derivable from Table 1; both must be corrected before the benchmark claims can be evaluated.","rationale":"I read the paper in good faith and find the proposed task and method plausible: the idea of using class-level text prompts to guide camouflaged object detection is reasonable, and the architecture (CPG + CGD) is systematically designed. The central claim, however, depends on the correctness of the benchmark numbers and the summary statistic. The dataset inconsistency between Sections 3 and 5.1 is a clear factual error that directly affects the reproducibility of the new dataset and the evaluation splits. The '5.5% better' claim is also a concrete numerical assertion that a careful reader cannot verify from Table 1; my own quick calculation of S-measure improvements yields about 1.3% on average, so the claim is at best undefined and at worst inflated. Both issues are fixable, so a conditional acceptance is appropriate, but the authors must provide corrected numbers and a precise definition of the average improvement. I considered other potential concerns, such as the circularity of the Hard/Normal split in Table 4 (which uses CGNet's own predictions to define the subsets), but that is a secondary robustness analysis; the dataset and the headline average improvement are more load-bearing for the paper's main message. The reader's weakest assumption identified the same dataset statistics issue, so I agree with that assessment and recommend no change to the conditional verdict.","tokens_in":14760,"tokens_out":10957,"duration_ms":97969,"concrete_test":"Write a script that parses Table 1, computes for each dataset the best and second-best S-measure among non-CGNet methods, and the average relative improvement of CGNet over the second-best; then cross-check the CamoClass totals by summing the test sizes reported in Table 1 and comparing with the 4,040 / 8,615 numbers in Section 5.1 and the 10,523 in Section 3. If the average improvement is not 5.5% or the test set total is not 6,473, the paper's claims must be revised.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing concern is that the paper's central quantitative claims are not internally consistent. Section 3 states CamoClass has 10,523 samples; Section 5.1 says the training set has 4,040 and the test set has 8,615, summing to 12,655. The per-dataset test sizes in Table 1 (250+2026+4121+76) sum to 6,473, so the '8,615' is off by 2,142. The standard training sets for CAMO and COD10K sum to 4,040, giving a total of 10,513, ten more than the stated 10,523. If the test split is not exactly the union of the four standard test sets, the benchmark comparisons and the Cseen/Cunseen analysis in Table 2 lose their reference. Furthermore, the claim in Section 5.2 that 'CGNet are on average 5.5% better than the second best method' is not supported by the numbers in Table 1: averaging the relative S-measure gains over the four datasets gives about 1.1–1.5% per dataset, and even averaging over all metrics does not obviously reach 5.5% without a precise definition. These two issues are concrete, checkable, and currently unresolved, so the paper's headline results rest on unverified statistics.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces class-guided camouflaged object detection (CGCOD), a new task in which class-level textual information is used to guide camouflaged object detection. To support this task, the authors construct CamoClass, a dataset assembled from CAMO, COD10K, NC4K, and CHAMELEON with class annotations added. They also propose CGNet, a framework consisting of a plug-and-play class prompt generator (CPG) built on frozen CLIP features and a class-guided detector (CGD), together with the CamoClass benchmark. The main experiments compare CGNet and several existing COD/RIS methods with and without the CPG on four test sets, and also report Cseen/Cunseen and hard/normal subset analyses.","tokens_in":15044,"tokens_out":6941,"duration_ms":59067,"significance":"If the numbers are correct, the paper makes a useful contribution by adding a class-guided setting to COD, providing a class-annotated dataset, and showing that a plug-and-play text prompt module can improve existing detectors. The evaluation is conducted on external benchmarks, the training/test separation follows the standard COD paradigm, and the code is publicly linked. However, the quantitative foundation currently has unresolved inconsistencies: the dataset totals in Section 3 and Section 5.1 do not match each other or the sums in Tables 1 and 2, and the headline 'average 5.5%' improvement is not reproducible from Table 1. These issues must be fixed before the benchmark claims can be fully trusted.","major_comments":[{"comment":"The dataset statistics are internally inconsistent. Section 3 states that CamoClass contains 10,523 samples, while Section 5.1 reports 4,040 training samples and 8,615 test samples, which sum to 12,655. Table 1 lists test-set sizes of 250 + 2,026 + 4,121 + 76 = 6,473, and Table 2 uses 6,109 + 364 = 6,473. The number 8,615 therefore does not equal the union of the four test sets, and 10,523 does not equal 4,040 + 6,473 = 10,513. Please correct the counts and specify exactly which images belong to CamoClass, how the 44 multi-class samples are counted, and how the train/test split is formed; until this is resolved, the benchmark comparisons and the Cseen/Cunseen analysis rest on unverified statistics.","section":"§3, §5.1, Tables 1 and 2"},{"comment":"The claim that 'CGNet are on average 5.5% better than the second best method' is not derivable from Table 1. Neither 'second best' nor 'on average' is defined in the text or the table. For example, averaging the relative S-measure improvements over the four datasets gives roughly 1.1–1.6% per dataset, not 5.5%. Please define the reference method, the metric, whether the percentage is absolute or relative, and show the arithmetic that produces 5.5% from Table 1; otherwise the headline quantitative claim is not verifiable.","section":"§5.2"},{"comment":"The reported plug-and-play improvements of the CPG are not traceable from Table 1. The text states that CPG improves SINet by 19.73%, PFNet by 12.68%, BGNet by 13.3%, and FSPNet by 4.43%. Under an obvious interpretation, such as averaging S-measure gains across the four datasets, SINet+ improves over SINet by about 7.8%, not 19.73%. Please specify the metric and the exact computation for each percentage, or revise the claim so that it can be checked against the table.","section":"§5.2"}],"minor_comments":[{"comment":"The manuscript text contains the local file path 'file:/E:/project/RefCOD-main/number_of_samples_by_class.png' where Figure 2 should be; the figure does not render in the PDF and must be replaced with an embedded image.","section":"§3, Figure 2"},{"comment":"The header 'Cuneen' should be 'Cunseen', and in Section 5.2 'FPSNet' should be 'FSPNet' for consistency with the method name used elsewhere.","section":"Table 2"},{"comment":"There are typos and grammatical errors, including 'tow parts' for 'two parts', 'feathers' for 'features', and 'traditional employ the traditional train paradigm' in Section 5.1; these should be corrected.","section":"§4"},{"comment":"The Hard/Normal split is defined by CGNet's own S-measure scores (Sm >= 0.9). Since the difficulty label is derived from the method being evaluated, this introduces a potential bias when comparing methods on the resulting subsets. Please justify this choice or replace it with a difficulty criterion independent of the evaluated method.","section":"§5.3, Table 4"},{"comment":"The criteria for the Cseen/Cunseen division are deferred to supplementary material. Please include the exact definition in the main text, as this split is used in Table 2 and is important for the generalization claims.","section":"§3"}],"recommendation":"major_revision","confidential_remarks":"The central idea is plausible and the experiments are extensive, but the manuscript's own numbers are not self-consistent. I would ask the authors to provide the exact dataset file counts and the arithmetic behind the 5.5% and per-detector improvement claims before the next round. The broken figure path suggests the PDF was compiled from an incomplete project directory; this should be checked at resubmission."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"CGCOD proposes a new task, class-guided camouflaged object detection; a dataset, CamoClass, with class annotations assembled from CAMO, COD10K, NC4K, and CHAMELEON; and CGNet, which uses frozen CLIP text and image features to generate a class prompt that guides a detector. The task is a reasonable extension of COD, and while text-guided COD already exists in ACUMEN and VSCode, the dedicated class-label setting and the plug-and-play framing are a real, incremental contribution. The paper ships code, uses standard benchmarks, and reports extensive ablations. Credit where due: the CPG plus class labels appears to improve several known detectors by noticeable margins, and the architecture is straightforward to reproduce.\n\nThe quantitative foundation needs work. The dataset counts do not add up: Section 3 says CamoClass has 10,523 samples; Section 5.1 reports 4,040 training and 8,615 test, summing to 12,655. The four test sets in Table 1 sum to 6,473, and the Cseen/Cunseen split in Table 2 also sums to 6,473 (6,109 + 364). The 8,615 figure is inconsistent with the tables, and the total differs from 4,040 + 6,473 by ten. These are the numbers that make the benchmark meaningful, so they need correction and explanation.\n\nThe \"5.5% better than second best\" claim in Section 5.2 is not derivable from Table 1. Averaging relative S-measure gains gives roughly 1-2%; the average might reach 5.5% under some metric definition, but the authors need to state what they averaged. As written, it is unsupported.\n\nMore serious: Table 4 splits test samples into Hard and Normal based on CGNet's own S-measure predictions. Using the method under evaluation to define difficulty subsets makes that comparison circular. The table should be redone with an independent difficulty criterion.\n\nOne more soft spot: the \"plug-and-play\" experiments retrain base detectors with class text, a frozen CLIP encoder, and the CPG. Gains over the original detectors therefore mix in extra supervision and model capacity; they do not isolate the CPG. The authors should disentangle these factors.\n\nNone of this makes me doubt the core idea. The paper is not circular in the sense of defining the task by the answer; class labels are used the way the task definition says they should be. The architecture is coherent. The soft spots are in reporting and evaluation design, not in the whole approach. I would send this to peer review with a strong request to fix the numbers, define the average, and replace the Table 4 protocol. If those are not corrected, the benchmark claims cannot be trusted. For a reading group, it is a maybe; I would hold off citing until the corrected version appears.","headline":"A useful incremental contribution with a real dataset, but the benchmark claims rest on inconsistent dataset counts and a circular hard/normal split.","tokens_in":15565,"tokens_out":3244,"would_cite":false,"duration_ms":32876,"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":"CGNet uses class-name text prompts to steer attention, claiming consistent state-of-the-art results on four camouflaged-object detection benchmarks and reporting that the prompt generator lifts existing detectors by 4.4–19.7%.","keywords":["class-guided camouflaged object detection","CamoClass dataset","text-guided segmentation","CLIP prompt","class prompt generator","multi-modal COD","plug-and-play module","zero-shot generalization"],"falsifier":"Recompute the CamoClass sample counts and the four test sets' sizes from the released data; if the totals cannot be reconciled with 10,523 images or with the reported train/test split, the benchmark numbers in Table 1 cannot be reproduced. A second check is to rerun CGNet with class prompts replaced by a generic word ('object') or a wrong class: if accuracy does not drop on the hard subset, the class-text guidance is not the active ingredient.","tokens_in":14577,"feed_emoji":"🎯","tokens_out":5892,"duration_ms":54012,"temperature":0.7,"pith_summary":"The paper proposes a new task, class-guided camouflaged object detection, in which the model receives not just an image but also the class name of the camouflaged object it should find. The authors argue that pure visual COD is unstable because camouflaged objects are defined by semantic difference, not visual contrast, and that explicit class text supplies the missing cue. To support this, they build CamoClass, a 10,523-image dataset with class labels, and design CGNet, a detector whose class prompt generator turns CLIP text embeddings into features that steer spatial and channel attention. The paper claims CGNet beats all compared COD and referring-segmentation methods on CAMO, COD10K, NC4K, and CHAMELEON, by an average 5.5% over the second-best, and that the prompt generator plugs into existing detectors and improves them.","feed_headline":"Class text beats visual-only COD on four benchmarks","feed_subtitle":"CGNet's class prompts report 5.5 points over the second-best detector and lift existing COD models when plugged in.","key_machinery":"The class prompt generator (CPG) is the load-bearing component: it combines a frozen CLIP text encoder, multi-level visual features from a frozen CLIP visual encoder, a cross-modal multi-head attention module that aligns text tokens with pixels, and a multi-level visual collaboration module (MVCM) that progressively refines the fused features. The class-guided detector (CGD) then uses a semantics consistency module (SCM) that applies the prompt feature as an attention guide in two stages—spatial localization followed by channel refinement—to keep the network focused on class-relevant foreground.","core_discovery":"The central claim is that class-level textual knowledge, injected through a plug-and-play class prompt generator, resolves the perceptual ambiguity that limits visual-only COD models. In CGNet, a frozen CLIP text encoder and a frozen CLIP visual encoder feed a cross-modal attention module that produces class prompt features; these features are fused with multi-level visual features and then guide a semantics consistency module that first localizes class-relevant regions spatially and then refines channel information. The paper reports consistent state-of-the-art results on four datasets and shows that appending the class prompt generator raises existing COD detectors' scores (for example, 19.73% for SINet, 13.3% for BGNet, 12.68% for PFNet, and 4.43% for FSPNet), attributing the gains to the model's ability to suppress false positives and to recover occluded or structurally degraded objects.","pith_inferences":["If the class-guidance mechanism is as transferable as claimed, the same CPG idea could be applied to other ambiguous segmentation domains, such as polyp boundaries in medical images or crop/weed separation in agriculture, where the semantic class is known but visual contrast is low.","The paper's split of test images into 'hard' and 'normal' based on its own detector's scores (Sm >= 0.9) is self-referential; a cleaner test would use human-rated camouflage difficulty, which would make the claimed benefit on hard scenes more independent of the model being evaluated.","The reported dataset arithmetic (10,523 samples versus 4,040 training plus 8,615 test) needs to be reconciled; if the true test set is 6,473 images, the average 5.5% margin could shift when recomputed on a consistent split.","A direct next experiment would be to replace class prompts with generic or incorrect class names; if CGNet's advantage persists, the performance gain is coming from the extra training data and architecture, not from class semantics."],"forward_implications":["If class guidance is the missing cue, COD becomes a multimodal task, and future benchmarks would need class annotations to stay comparable with this new paradigm.","Existing COD detectors do not need redesign: the CPG can be appended to them and retrained on CamoClass, yielding reported gains of 4.4% to 19.7% on the tested baselines.","The seen/unseen split of CamoClass (Cseen and Cunseen) provides a way to measure zero-shot generalization of COD models, and the paper reports CGNet leads on both splits, especially on Cunseen.","The class prompt helps most in hard cases—multi-object scenes, occlusion, edge-similar backgrounds, and small objects—where the paper shows visual-only methods fail.","The proposed loss and multi-level supervision tie the prompt generator, the detector, and the final prediction into one trainable pipeline, making the class-guidance effect a direct outcome of optimization rather than a post-hoc refinement."],"supporting_citations":[{"why":"Supplies the frozen CLIP text and visual encoders from which CGNet builds class prompt features; the paper's core mechanism depends on CLIP's text–image alignment.","marker":"[Radford et al., 2021]"},{"why":"Provides the SINet baseline and the camouflaged object detection problem formulation that CGNet extends; also one of the detectors used in the plug-and-play experiments.","marker":"[Fan et al., 2020a]"},{"why":"Provides the COD10K dataset used for training and evaluation, plus the standard evaluation protocol.","marker":"[Fan et al., 2022]"},{"why":"Provides the CAMO dataset used both for training and as a test benchmark.","marker":"[Le et al., 2019]"},{"why":"Provides the NC4K test benchmark.","marker":"[Lv et al., 2021]"},{"why":"Provides the CHAMELEON test benchmark.","marker":"[Przemysław et al., 2018]"},{"why":"The prior visual–textual COD method (ACUMEN) that this paper compares against and whose coarse-grained text approach it argues against with class-level labels.","marker":"[Zhang et al., 2025]"},{"why":"Provides the PVT backbone of the class-guided detector, the architecture that the pluggable prompt generator is attached to.","marker":"[Wang et al., 2021]"}],"fun_headline_variants":["Class prompts push COD detectors to new highs","Textual cues sharpen camouflaged object detection","Class-guided COD gains 5.5 points over SOTA","Plug-in class prompts boost COD across the board","CGNet: Text knowledge resolves COD ambiguity"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper's benchmark and generalization claims rest on the CamoClass dataset being complete and internally consistent, but the reported totals do not match (10,523 samples versus 4,040 training plus 8,615 test), so the exact composition and splits are the load-bearing premise.","fun_headline_variants_meta":{"raw":{"variants":["Class prompts push COD detectors to new highs","Textual cues sharpen camouflaged object detection","Class-guided COD gains 5.5 points over SOTA","Plug-in class prompts boost COD across the board","CGNet: Text knowledge resolves COD ambiguity"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000211,"raw_usage":{"total_tokens":1405,"prompt_tokens":927,"completion_tokens":478,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":543,"completion_tokens_details":{"reasoning_tokens":406}},"tokens_in":543,"tokens_out":478,"duration_ms":5934,"temperature":1.0,"reasoning_tokens":406,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T00:57:22.538675+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Recompute the CamoClass sample counts and the four test sets' sizes from the released data; if the totals cannot be reconciled with 10,523 images or with the reported train/test split, the benchmark numbers in Table 1 cannot be reproduced. A second check is to rerun CGNet with class prompts replaced by a generic word ('object') or a wrong class: if accuracy does not drop on the hard subset, the class-text guidance is not the active ingredient.","supporting_citations":[{"cited_title":"Concealed object detection","cited_arxiv_id":null,"evidence_quote":"Provides the COD10K dataset used for training and evaluation, plus the standard evaluation protocol."},{"cited_title":"Open-vocabulary camouflaged object segmentation","cited_arxiv_id":null,"evidence_quote":"The prior visual–textual COD method (ACUMEN) that this paper compares against and whose coarse-grained text approach it argues against with class-level labels."},{"cited_title":"Pyramid vision transformer: A versatile backbone for dense prediction without convolutions","cited_arxiv_id":null,"evidence_quote":"Provides the PVT backbone of the class-guided detector, the architecture that the pluggable prompt generator is attached to."}],"review_version":1}