{"id":"1f021e09-38da-4bcf-8a8b-c97b8509eb36","arxiv_id":"2412.10943","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A new dataset, model, and metric extend salient and camouflaged object detection to unconstrained scenes where the two object types can coexist.","lead":"The authors present USC12K, a 12,000-image dataset in which scenes may contain salient objects, camouflaged objects, both, or neither, plus a model named USCNet that explicitly models the relationship between saliency and camouflage. The paper also introduces a confusion metric, CSCS, and reports that training on USC12K sharply reduces cross-task misclassification errors.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Scene C's ground truth is partly SAM-generated while USCNet is SAM-based, and no inter-annotator agreement is reported; the SOTA claim may be circular.","rationale":"The reader's conditional verdict is correct, and the annotation-quality concern is the most load-bearing assumption. The paper itself provides the key evidence for the concern: Scene C masks are initialized by SAM, and the proposed method is built on SAM/SAM2. The appendix describes a voting and refinement process but gives no quantitative reliability measures, so the novel part of USC12K is unverified. I do not see an internal inconsistency in the CSCS formula or the ARM module large enough to overturn the conditional verdict; the ablation study and generalization tables are supporting evidence, but they all inherit the same dataset labels. The requested release of code, dataset, and QA statistics would address the concern. My concrete test isolates the circularity by re-annotating without SAM and checking whether USCNet's advantage is robust to an independent ground truth. If the test passes, the central claim is substantially strengthened; if it fails, the dataset and SOTA conclusion need revision. Therefore the paper should remain conditional, not accepted as-is and not rejected outright.","tokens_in":23022,"tokens_out":11887,"duration_ms":115030,"concrete_test":"Select a random sample of 200 Scene C images and 100 Scene D images from USC12K. Have 5 independent annotators re-annotate saliency and camouflage masks using a non-SAM interactive tool (plain brush/polygon annotation, no SAM assistance). Compute (i) pixel-level IoU and Fleiss' kappa between annotators and between annotators and the released labels; (ii) evaluate USCNet and the strongest non-SAM competitor (e.g., ICEG or CamoFormer) on the same images using the original USC12K training set. If annotator agreement is high (e.g., mean pairwise IoU above 0.85) and USCNet's mIoU/CSCS margin over the competitor is preserved on the non-SAM labels, the circularity concern is resolved; if the margin shrinks or agreement is low, the dataset labels and the leaderboard need revision before the SOTA claim is accepted.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"USC12K's central novelty is Scene C (3,000 images containing both salient and camouflaged objects), and the claim that USCNet achieves SOTA across all scenes depends on these labels being an independent ground truth for unconstrained scenes. In Section 3.2, the authors state that for Scene C they 'use SAM for coarse annotation, followed by manual refinement,' and Appendix Section 5 describes mask annotation with the ISAT tool, which supports SAM semi-automatic labeling. USCNet itself is built on the SAM/SAM2 encoder and a frozen SAM mask decoder. If the manual refinement was light, the benchmark's target masks inherit SAM's segmentation priors, so USCNet is evaluated against ground truth produced by the same model family it uses, while competitors without SAM do not receive that benefit. This is a potential circularity, not just random annotation noise. In addition, no inter-annotator agreement statistics (kappa, pixel IoU, or per-scene disagreement rates) are reported for the 2,617 web-collected Scene C images or the 1,436 Scene D images. The Appendix describes voting by 7 observers and refinement by 3 observers but never quantifies agreement. Without an independent check, the leaderboard and the 'SOTA across all scenes' conclusion could reflect SAM-derived annotation bias and annotator subjectivity rather than real unconstrained detection ability.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper claims that existing SOD/COD datasets and models enforce a mutual-exclusivity constraint on salient and camouflaged objects, which causes cross-task misinterpretation. To address this, the authors introduce USC12K, a 12,000-image dataset with four scene types covering all logical combinations of salient and camouflaged objects (salient-only, camouflaged-only, both, neither), propose a SAM2-based model USCNet with an Attribute Relation Modeling (ARM) module that uses inter-sample and intra-sample prompt queries, and design a new evaluation metric, the Camouflage-Saliency Confusion Score (CSCS). They retrain 21 SOD/COD/unified methods on USC12K, report that USCNet achieves state-of-the-art performance across all scenes and metrics, and demonstrate that training on USC12K reduces cross-task false positives.","tokens_in":23294,"tokens_out":10847,"duration_ms":92907,"significance":"If validated, USC12K would be a valuable new benchmark for unconstrained salient/camouflaged object detection, and the CSCS metric addresses a real gap in evaluating the confusion between these two attributes. The ARM module is a sensible approach for explicitly modeling attribute relationships, and the authors provide a fairly extensive comparison, including generalization results on six standard SOD/COD datasets. The paper also ships promising reproducibility artifacts: it promises to release code and data, and the benchmark protocol is described in detail. The main risks are that the Scene C ground truth is partly SAM-derived without quantified annotation agreement, and the scene definitions are internally inconsistent; these issues must be resolved before the benchmark and SOTA claims can be fully trusted.","major_comments":[{"comment":"Section 3.1 defines Scene A as containing only salient objects and Scene B as containing only camouflaged objects, but Figure 3 captions these scenes in the opposite order, and the numbers in Table 3 (e.g., high IoUC values in Scene A and high IoUS values in Scene B) are only interpretable under the swapped definitions. This internal inconsistency must be resolved for the benchmark results to be meaningful.","section":"Section 3.1 vs. Figure 3 / Table 3"},{"comment":"The authors state that Scene C annotations are produced using SAM for coarse annotation followed by manual refinement, and the appendix describes the ISAT tool with SAM semi-automatic labeling. No inter-annotator agreement statistics or quantitative measure of the refinement effort are provided. Because USCNet uses SAM2 as its backbone and a frozen SAM mask decoder, the Scene C benchmark may be biased in favor of SAM-based models. Please provide independent validation, such as pixel-level agreement among annotators, the distribution of manual corrections, and evaluation on a subset with fully manual masks.","section":"Section 3.2 and Appendix §5"},{"comment":"All quantitative comparisons report single-run point estimates without variance, and several margins over the second-best method are small (e.g., Table 3, Scene A IoUS: USCNet 79.70 vs. SAM2-Adapter 78.75). Without multiple seeds or significance tests, the claim of state-of-the-art performance across all scenes is not robust.","section":"Tables 3, 5, 7–10"},{"comment":"The benchmark protocol retools every conventional SOD/COD model's output layer into a three-class softmax and modifies the training procedure for unified models (using two copies of the dataset for VSCode and EVP). These adaptations may disadvantage the competitors relative to USCNet, so the authors should justify their neutrality or supplement the benchmark with baselines that use the models' native binary output heads plus a post-hoc background class.","section":"Appendix §4"}],"minor_comments":[{"comment":"The CSCS denominator can vanish when no pixels are predicted as salient or as camouflaged; please specify the convention for such degenerate cases.","section":"Eq. (6) / Eq. (7)"},{"comment":"The text says 21 methods are compared, but the enumerated list omits PGNet, which appears in Table 3; please ensure the count and the list are consistent.","section":"Section 5"},{"comment":"The caption of Figure 3 reverses the definitions of Scene A and Scene B relative to Section 3.1; beyond the major comment, please ensure all captions use consistent scene definitions.","section":"Figure 3"},{"comment":"The term 'inter-sample' for the Inter-SPQ queries is potentially misleading, because the queries are fixed at inference and do not depend on other samples; consider renaming them 'global' or 'dataset-level' queries.","section":"Section 4.2"},{"comment":"The column labeled 'Update' is unexplained; please clarify whether it indicates that all models are retrained on USC12K or some other setting.","section":"Table 3"}],"recommendation":"major_revision","confidential_remarks":"The scene definition inconsistency in Section 3.1 versus Figure 3 and Table 3 is a load-bearing issue that must be fixed. The annotation circularity concern for Scene C is real: the paper's own description indicates SAM involvement in ground-truth generation, and the absence of agreement statistics weakens the benchmark's validity. These are fixable with additional validation and clarification, so I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This is a solid dataset paper, and the dataset is the real contribution. USC12K is the first benchmark, as far as I know, that jointly annotates salient and camouflaged objects in the same image and includes scenes with neither. That fills a genuine gap: existing SOD/COD datasets enforce mutual exclusivity, and the paper shows concretely (Table 1 vs Table 4) that training on USC12K reduces the cross-task misdetection. The CSCS metric is a sensible addition, and the ARM module with inter-/intra-sample prompt queries is a reasonable model idea, even if it's not the main event.\n\nWhat the paper does well: the benchmark tables are thorough, with 21 competitors all retrained on the same splits, and the generalization results on DUTS/HKU-IS/COD10K/NC4K show consistent gains. The ablation in Table 6 is clean and isolates the ARM components. The appendix confusion matrix is a nice diagnostic.\n\nSoft spots, in proportion. The stress-test concern about Scene C being SAM-generated while USCNet is SAM-based is real but not fatal. The paper does describe a voting process and manual refinement by 3 observers; it's not just SAM's raw output. Still, the lack of any inter-annotator agreement statistic (kappa, pixel IoU, per-scene disagreement) is the bigger gap for a dataset paper. The description 'use SAM for coarse annotation, followed by manual refinement' needs quantification. I'd ask the authors to report how much refinement was needed and provide agreement numbers.\n\nThe comparison protocol modifies the output layers of all competitors to produce three classes; that's defensible given the task, but it means the leaderboard doesn't reflect the original model designs. Single-run results without error bars are common in this subfield, but with a new benchmark they matter more.\n\nThe biggest practical issue: code and dataset are not released. Without the artifacts, the SOTA claim can't be verified and the benchmark can't be used. That's a standard conditional.\n\nBottom line: this deserves peer review. The dataset alone is significant for the SOD/COD community. I'd accept the paper pending release of dataset and code, and require annotation-agreement statistics and a statement on SAM's role in annotation. The SOTA claims are plausible, so I'd let the referees check them against the released data.","headline":"A genuinely useful four-scene dataset with a plausible SOTA method; the main gaps are missing annotation agreement stats and unreleased artifacts, not the SAM-annotation circularity that dominates the stress-test note.","tokens_in":23823,"tokens_out":3343,"would_cite":true,"duration_ms":30045,"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":"This paper claims that SOD/COD models fail in unconstrained scenes because datasets assume scenes contain either salient or camouflaged objects but never both, and that a new dataset (USC12K), a new model (USCNet), and a new metric (CSCS)…","keywords":["salient object detection","camouflaged object detection","unconstrained scenes","USC12K dataset","attribute relationship modeling","Segment Anything Model","CSCS metric"],"falsifier":"One concrete experiment that would settle the central claim is to re-annotate a stratified random sample of the 3,000 Scene C images and the 1,436 internet Scene D images with independent expert annotators and compute per-pixel inter-annotator agreement (e.g., Cohen's kappa) against the released labels; if agreement is low, or if the original labels are systematically biased toward 'salient' or 'camouflaged' in ambiguous cases, the USC12K ground truth is unstable and the reported CSCS ranking of USCNet over SAM2-Adapter could be an artifact of that label noise rather than a genuine modeling gain.","tokens_in":22844,"feed_emoji":"🎯","tokens_out":7077,"duration_ms":57673,"temperature":0.7,"pith_summary":"The paper argues that salient object detection (SOD) and camouflaged object detection (COD) fail in unconstrained real scenes for two connected reasons: existing datasets force a scene to contain either salient or camouflaged objects, never both and never neither, and existing models never explicitly model the relationship between the two attributes. To break this assumption, the authors construct USC12K, a 12,000-image dataset with four balanced scene types covering every logical combination of salient/camouflaged presence, including 3,000 images containing both and 3,000 containing neither. They also propose USCNet, a unified SAM-based network whose Attribute Relation Modeling (ARM) module uses inter-sample and intra-sample prompt queries to separate salient from camouflaged features, and a new metric, CSCS, that directly measures how often a model swaps the two attributes. If the paper is right, training on USC12K eliminates the label-conflict bias that makes SOD models call camouflaged objects salient and vice versa, and the reported state-of-the-art performance across all scenes shows the benefit of explicit attribute-relationship modeling.","feed_headline":"Salient or camouflaged? A new 12K dataset says: both, or neither","feed_subtitle":"USC12K adds scenes where both (or neither) appear; USCNet models the relation; new metric CSCS measures confusion.","key_machinery":"The load-bearing pieces are threefold. First, USC12K: a 12,000-image benchmark with four balanced scene types (only salient, only camouflaged, both, neither) and labels for the three attributes of saliency, camouflage, and background, built from existing DUTS, HKU-IS, COD10K, and CAMO datasets plus newly collected internet and underwater images. Second, the Attribute Relation Modeling (ARM) module inside USCNet: it generates three attribute prompts (saliency, camouflage, background) by combining Inter-SPQ (learnable cross-sample queries) and Intra-SPQ (per-sample queries computed as $[Q_{Sa}, Q_{Ca}, Q_{Ba}] = \\mathrm{Linear}(\\sigma(\\Phi_{AH}(F)) \\otimes F)$), then applies self-attention and query-to-image attention to produce the prompts that feed a frozen SAM mask decoder; this is the mechanism that explicitly models how the two attributes relate both across samples and within a single image. Third, the Camouflage-Saliency Confusion Score, $\\mathrm{CSCS} = \\frac{1}{2}\\left(\\frac{P_{CS}}{P_{BS}+P_{SS}+P_{CS}} + \\frac{P_{SC}}{P_{BC}+P_{SC}+P_{CC}}\\right)$, which quantifies how often camouflaged pixels are predicted as salient and vice versa, a quantity that existing foreground/background metrics such as mIoU and weighted F-measure do not isolate.","core_discovery":"On its own terms, the central discovery is that the mutual-exclusivity annotation paradigm of existing SOD/COD datasets—labeling a scene either salient or camouflaged, but never both and never neither—systematically corrupts both single-task and unified detectors, and that this can be fixed by two complementary contributions. First, a dataset, USC12K, whose four balanced scenes (A: only salient, B: only camouflaged, C: both, D: neither) carry full pixel-level masks for all three attributes and whose Scene C contains both salient and camouflaged objects in the same image. Second, a model, USCNet, that explicitly models the relationship between the two attributes through an Attribute Relation Modeling module using Inter-Sample Prompt Query (learnable, sample-invariant queries that capture generic attribute difference) and Intra-Sample Prompt Query (per-image queries derived from an attention-weighted feature map that capture within-sample context), built on a SAM/SAM2 encoder with adapters and a frozen mask decoder. The paper reports that USCNet achieves state-of-the-art performance across all four scenes and all metrics on the USC12K benchmark, and that training on USC12K sharply reduces the cross-task misinterpretation scores shown in Table 1, for example a SOD model's F_beta score on COD10K dropping from 0.6384 to 0.0146.","pith_inferences":["Editorial inference: the mutual-exclusivity bias probably affects other binary foreground tasks beyond SOD/COD, such as defect detection or anomaly segmentation, so the four-scene design of USC12K could be transferred to those domains to test whether models similarly confuse opposing attributes.","Editorial inference: the Scene D background-only images, which are often trivial for standard metrics, may be the most important test for real deployment; a detector that reports false positives on them would be unsafe, and the CSCS-style evaluation could be extended to measure background-foreground false alarms in addition to attribute confusion.","Editorial inference: a minimal attribute-switch experiment could be run with USCNet—keep the encoder and decoder fixed, swap only the saliency and camouflage attribute prompts; if the prompts truly carry attribute information, the same image should produce the opposite mask, which would isolate whether the model disentangles attributes rather than learning a single foreground detector.","Editorial inference: if the annotation-quality assumption holds, retraining existing SOD and COD models on USC12K should produce larger improvements than architecture changes alone, a prediction that the paper's Table 4 already supports for a few baselines and that could be tested on a broader set of models."],"forward_implications":["If USC12K is adopted as a training set, SOD and COD models should no longer exhibit the cross-task misinterpretation shown in Table 1; the paper reports that retraining baselines on USC12K drops their false-positive F_beta scores on the opposite task from values like 0.6384 to below 0.1.","In scenes containing both salient and camouflaged objects (Scene C), explicit inter- and intra-sample attribute modeling yields the largest gains; the ablation shows that both Intra-SPQ and Inter-SPQ improve IoU and mIoU and lower CSCS, with Intra-SPQ contributing more.","Because USCNet attaches the ARM module to a frozen SAM2 mask decoder with only 4.04 million tunable parameters, the claimed state-of-the-art results come with parameter-efficient fine-tuning, suggesting the approach can be ported to other large vision backbones.","The proposed CSCS metric provides a direct way to compare detectors on salient-versus-camouflaged discrimination, complementing standard foreground/background metrics and giving the community a concrete target for reducing attribute confusion in unconstrained scenes."],"supporting_citations":[{"why":"Segment Anything Model (SAM); the image encoder and frozen mask decoder that USCNet builds on, providing scene generalization and the prompt-driven mask prediction architecture.","marker":"[23]"},{"why":"SAM2; used as the base model (hiera-large) for USCNet in the main experiments, and its frozen mask decoder is compared against in ablations.","marker":"[58]"},{"why":"SAM-Adapter; supplies the parameter-efficient adapter strategy that USCNet uses to inject SOD/COD-specific knowledge into the SAM encoder.","marker":"[2]"},{"why":"COD10K; one of the main camouflaged-object datasets and the source of most Scene B and Scene D images, as well as existing category labels for camouflaged objects.","marker":"[11]"},{"why":"CAMO; a camouflaged-object dataset used as a source of Scene B images and in the benchmark comparison.","marker":"[24]"},{"why":"NC4K; a camouflaged-object dataset that provides some Scene C images with both salient and camouflaged objects and is used in generalization tests.","marker":"[41]"},{"why":"DUTS; a salient-object dataset that is the primary source of Scene A images in USC12K and a standard benchmark for the SOD generalization tests.","marker":"[66]"},{"why":"HKU-IS; a salient-object dataset used as an additional source of Scene A images and a benchmark for generalization.","marker":"[27]"},{"why":"VSCode; the strongest unified SOD/COD baseline that USCNet compares against, using 2D prompt learning to distinguish saliency and camouflage, and whose cross-task misinterpretation measurements motivate the paper's central claim.","marker":"[40]"}],"fun_headline_variants":["Both or neither: 12K dataset breaks the salient/camouflaged either-or","New dataset puts salient and camouflaged in the same image—USCNet tells them apart","SOD and COD models clash—USC12K shows why and USCNet resolves it","USCNet: one model that separates salient from camouflaged in any scene","Dataset with both/neither scenes: models learn to stop mislabeling camouflage"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The label quality of the newly internet-collected images (2,617 'both' and 1,436 'neither' images) is asserted through a voting and refinement process without reported inter-annotator agreement statistics; if those labels are noisy, systematically biased, or inconsistent with existing SOD/COD label conventions, the benchmark comparisons and the conclusion that USCNet achieves state-of-the-art results would not transfer to real unconstrained scenes.","fun_headline_variants_meta":{"raw":{"variants":["Both or neither: 12K dataset breaks the salient/camouflaged either-or","New dataset puts salient and camouflaged in the same image—USCNet tells them apart","SOD and COD models clash—USC12K shows why and USCNet resolves it","USCNet: one model that separates salient from camouflaged in any scene","Dataset with both/neither scenes: models learn to stop mislabeling camouflage"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000762,"raw_usage":{"total_tokens":3477,"prompt_tokens":1133,"completion_tokens":2344,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":749,"completion_tokens_details":{"reasoning_tokens":2231}},"tokens_in":749,"tokens_out":2344,"duration_ms":15352,"temperature":1.0,"reasoning_tokens":2231,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T15:27:27.665894+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"One concrete experiment that would settle the central claim is to re-annotate a stratified random sample of the 3,000 Scene C images and the 1,436 internet Scene D images with independent expert annotators and compute per-pixel inter-annotator agreement (e.g., Cohen's kappa) against the released labels; if agreement is low, or if the original labels are systematically biased toward 'salient' or 'camouflaged' in ambiguous cases, the USC12K ground truth is unstable and the reported CSCS ranking of USCNet over SAM2-Adapter could be an artifact of that label noise rather than a genuine modeling gain.","supporting_citations":[{"cited_title":"Segment any- thing","cited_arxiv_id":null,"evidence_quote":"Segment Anything Model (SAM); the image encoder and frozen mask decoder that USCNet builds on, providing scene generalization and the prompt-driven mask prediction architecture."},{"cited_title":"Sam 2: Seg- ment anything in images and videos","cited_arxiv_id":null,"evidence_quote":"SAM2; used as the base model (hiera-large) for USCNet in the main experiments, and its frozen mask decoder is compared against in ablations."},{"cited_title":"Anabranch network for camouflaged object segmentation","cited_arxiv_id":null,"evidence_quote":"CAMO; a camouflaged-object dataset used as a source of Scene B images and in the benchmark comparison."},{"cited_title":"Simultaneously localize, segment and rank the camouflaged objects","cited_arxiv_id":null,"evidence_quote":"NC4K; a camouflaged-object dataset that provides some Scene C images with both salient and camouflaged objects and is used in generalization tests."},{"cited_title":"Visual saliency based on multi- scale deep features","cited_arxiv_id":null,"evidence_quote":"HKU-IS; a salient-object dataset used as an additional source of Scene A images and a benchmark for generalization."},{"cited_title":"Vscode: General visual salient and camouflaged object de- tection with 2d prompt learning","cited_arxiv_id":null,"evidence_quote":"VSCode; the strongest unified SOD/COD baseline that USCNet compares against, using 2D prompt learning to distinguish saliency and camouflage, and whose cross-task misinterpretation measurements motivate the paper's central claim."}],"review_version":1}