{"id":"2b454d92-568f-4b81-a0d9-22d51acf1e08","arxiv_id":"2509.03786","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"SLENet, a SAM2-adapter network with gamma-asymmetric enhancement and localization guidance, reportedly beats prior methods on underwater camouflaged object detection, and DeepCamo is introduced as a new benchmark.","lead":"This paper introduces DeepCamo, a benchmark of 2,493 underwater camouflaged images, and SLENet, a deep network built on a frozen SAM2 encoder with adapters, a gamma-asymmetric enhancement module, a localization guidance branch, and a multi-scale supervised decoder. SLENet reports top scores on DeepCamo and three camouflage benchmarks, with its largest gains on its own dataset.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table 2's DeepCamo comparison is not head-to-head: baselines are apparently evaluated zero-shot with pretrained weights while SLENet trains on DeepCamo-train, so the reported SOTA margin is not a fair architecture comparison.","rationale":"The reader's weakest assumption is the load-bearing one: Table 2's DeepCamo-Test comparison is asymmetric because SLENet is trained on DeepCamo-train while baselines are, per Sec. 3.3 and the default reading of Table 2, evaluated with public pretrained weights. My stress-test adds two precise observations that make the concern more concrete: (1) Table 1 vs Table 2 shows large jumps for identical baselines on two DeepCamo splits, which needs an explanation about split difficulty or training protocol; and (2) the universal 'all metrics' claim is internally contradicted by CAMO-Test Fwβ, where SLENet (0.860) is below three baselines (0.861). These are correctness risks, not stylistic objections. They do not overturn the paper's possible value as a dataset/benchmark, and the architecture is plausible; hence the verdict remains conditional rather than reject. No ad hominem is intended; the issue is the undisclosed/unfair evaluation protocol and the overbroad wording.","tokens_in":10219,"tokens_out":5302,"duration_ms":46353,"concrete_test":"Obtain the released checkpoints of the 10 baselines and run a controlled fine-tuning protocol on DeepCamo-train identical to SLENet's (352×352, batch 16, AdamW lr=5e-4 cosine, 100 epochs), then evaluate on DeepCamo-Test. If any baseline matches or beats SLENet on Sα, Eϕ, Fwβ, or MAE, the unconditional SOTA claim fails. As a faster consistency check, ask the authors to disclose whether Table 2's baseline rows use pretrained weights or fine-tuned weights, and to recompute the same baselines on the released DeepCamo-full split so Table 1 and Table 2 use a common protocol. Also verify the single CAMO-Test Fwβ value: if 0.860 is correct, the sentence in Sec. 5.2 must be corrected.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim (Abstract; Sec. 5.2) is that SLENet 'consistently outperforms all competing ... across all evaluation metrics.' The most load-bearing support is Table 2 on DeepCamo-Test, but the protocol there is not described as a like-for-like training comparison. Sec. 3.3 states that benchmarked methods are evaluated with 'publicly available pretrained weights'; Table 2 never says whether those same baselines were fine-tuned on DeepCamo-train. SLENet is trained on DeepCamo-train for 100 epochs with adapters. Because DeepCamo-Test is from the same distribution as DeepCamo-train, a zero-shot baseline comparison conflates domain adaptation with architectural superiority. The inconsistency between Table 1 and Table 2 for the same baselines reinforces the problem: e.g., SINet Sα rises from 0.665 on DeepCamo-full to 0.745 on DeepCamo-Test, and SAM2-UNet from 0.741 to 0.859. If both columns use pretrained weights, the test split is far easier / not comparable to the benchmark subset; if baselines were fine-tuned for Table 2, that protocol is undisclosed. Independently, the 'all metrics' statement is already false in Table 2: on CAMO-Test, SLENet's Fwβ is 0.860 while FSNet, HitNet, and SAM2-UNet achieve 0.861. Either the comparison is asymmetric or the headline overstates the numbers; both need resolution before the SOTA claim can be accepted.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript defines underwater camouflaged object detection (UCOD), introduces the DeepCamo dataset (2,493 images, 16 marine species), and proposes SLENet, which combines a frozen SAM2 Hiera-L encoder with adapters, a Gamma-Asymmetric Enhancement (GAE) module, a Localization Guidance Branch (LGB), and a Multi-Scale Supervised Decoder (MSSD). The loss is a weighted BCE+IoU together with a separately supervised localization map whose weight decays linearly. The paper benchmarks 10 COD methods on DeepCamo and three public COD datasets and claims that SLENet consistently outperforms all competitors on all metrics.","tokens_in":10646,"tokens_out":4196,"duration_ms":42054,"significance":"If the results are supported, DeepCamo fills an actual gap in camouflaged object detection, and SLENet offers a plausible strong baseline for the UCOD setting. The paper has concrete strengths: it introduces a new benchmark with explicit attention to multi-object scenes and underwater degradation, evaluates on four datasets, uses multiple standard metrics, and reports module-wise ablations including sensitivity to the loss weight μ. The main weakness is that the headline SOTA claim is contradicted by the paper's own Table 2 in several cells, and the DeepCamo comparison is not transparently head-to-head because baselines are evaluated with pretrained weights while SLENet is trained on DeepCamo-train. These issues are correctable but currently load-bearing for the central claim.","major_comments":[{"comment":"The sentence 'SLENet consistently outperforms all competing CNN-based and Transformer-based methods across all evaluation metrics' is contradicted by Table 2. On CAMO-Test, FSNet and SAM2-UNet achieve Fwβ=0.861 while SLENet achieves 0.860. On CHAMELEON, HitNet achieves Fwβ=0.903 while SLENet achieves 0.896. The claim should be restricted to 'most metrics' or supported with paired significance tests and error bars.","section":"Sec. 5.2, Table 2"},{"comment":"The DeepCamo comparison in Table 2 is not shown to be a like-for-like training comparison. Sec. 3.3 states that baselines are evaluated with 'publicly available pretrained weights' on DeepCamo-full; for Table 2's DeepCamo-Test column the paper does not state whether the same baselines were fine-tuned on DeepCamo-train, whereas SLENet is explicitly trained on DeepCamo-train for 100 epochs. If the baselines are zero-shot, the reported margins on DeepCamo-Test conflate domain-transfer with architectural superiority, and the SOTA claim on the new dataset is not established. The paper should either fine-tune all baselines on DeepCamo-train under comparable schedules or clearly mark the DeepCamo numbers as zero-shot transfer and rephrase the claims.","section":"Sec. 3.3, Sec. 5.1, Table 2"},{"comment":"The relationship between DeepCamo-full (1,907 images), DeepCamo-train (1,931), and DeepCamo-test (562) is unspecified, and this creates an apparent internal inconsistency. For the same baseline methods, Table 1 and Table 2 report very different numbers: SINet Sα=0.665 on DeepCamo-full versus 0.745 on DeepCamo-Test; SAM2-UNet Sα=0.741 versus 0.859. If both columns use pretrained weights, the DeepCamo-Test split appears substantially easier, which undermines the claim that the new benchmark is uniformly challenging. If fine-tuning was applied for Table 2, that protocol must be disclosed. The authors should clarify how DeepCamo-full relates to the train/test split and reconcile the two tables.","section":"Sec. 3.2-3.3, Tables 1 and 2"}],"minor_comments":[{"comment":"The wording 'superior performance over SOTA methods' and 'consistently outperforms' should be softened to match the quantitative evidence, particularly after the Table 2 issues above are resolved.","section":"Abstract; Sec. 5.2"},{"comment":"The notation μ, epoch, and epochs is used without precise definition; state that μ is the initial localization-map loss weight and clarify the floor value 0.1 in the linear decay.","section":"Eq. (13), Sec. 4.5"},{"comment":"The 'weighted BCE' loss is said to use weights based on pixel contrast, but no formula or precise weighting rule is given. Please provide the exact definition so the loss is reproducible.","section":"Sec. 4.5"},{"comment":"The architecture figures are dense and some arrows are hard to read at normal print size; consider enlarging the figures or adding a table of module configurations.","section":"Fig. 2, Fig. 3"},{"comment":"The text states that μ=0 leads to a 'significant performance drop', but the corresponding row is omitted from Table 3b. Since the table otherwise reports μ values, including μ=0 would make the claim checkable.","section":"Sec. 5.3, Table 3b"},{"comment":"The paper does not state whether the DeepCamo dataset and code will be released. For a benchmark contribution, dataset accessibility is important and should be addressed.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The main risk is the fairness of the DeepCamo comparison. If the authors can either fine-tune the baselines on DeepCamo-train or explicitly present Table 2 as zero-shot transfer with appropriate qualifiers, the paper's empirical claims can be brought in line with the data. The CAMO/CHAMELEON discrepancies in Table 2 also need to be fixed in the text before the paper can be considered for publication. The dataset contribution is potentially valuable, but its release would substantially increase its impact."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Hi,\n\nQuick read of SLENet. The real contribution is the dataset. DeepCamo is a new benchmark with 2,493 underwater camouflaged images, and the benchmark study in Table 1 shows existing COD methods degrade sharply on it. That is a useful, credible piece of evidence.\n\nThe network is a competent mashup of known components: a frozen SAM2 encoder with adapters, cascaded asymmetric and dilated convolutions, channel/spatial attention, and SPADE-style affine modulation. Nothing conceptually new, but the combination works reasonably well on the data.\n\nThe soft spots are concentrated in Section 5.2 and Table 2. First, the DeepCamo-Test comparison is not head-to-head. The baselines appear to be evaluated with their pretrained weights (that's the protocol stated for the DeepCamo-full benchmark in Table 1), while SLENet is trained on DeepCamo-train. Since train and test come from the same distribution, the margins on DeepCamo-Test conflate domain adaptation with architecture. The paper never discloses whether the baselines were fine-tuned on DeepCamo-train. As it stands, the SOTA claim on DeepCamo is not supported. That is the load-bearing issue.\n\nSecond, the abstract and Section 5.2 say SLENet 'consistently outperforms all competing methods across all evaluation metrics.' That is simply false: on CAMO-Test, SLENet's weighted F-measure is 0.860, while FSNet, HitNet, and SAM2-UNet get 0.861. It's a small gap, but it is a direct contradiction.\n\nThird, the mu=0 ablation is omitted. The text claims a significant performance drop but leaves the number out 'to better illustrate the key trend.' That is selective reporting. Either report the number or don't claim the ablation.\n\nAlso, no code or dataset release is provided, which limits the benchmark's immediate reproducibility.\n\nIf the authors fix the comparison protocol, disclose the mu=0 result, and soften the overclaim, this is a useful paper for the COD/UCOD community. The dataset alone is worth having.\n\nRecommendation: send it to peer review, but require the fine-tuning protocol for baselines to be disclosed before acceptance.","headline":"DeepCamo is a genuinely useful new benchmark, but the paper's SOTA claim on it rests on an apples-to-oranges comparison and a false 'all metrics' statement.","tokens_in":11113,"tokens_out":2912,"would_cite":false,"duration_ms":27475,"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 defines underwater camouflaged object detection, releases the DeepCamo benchmark, and claims SLENet beats ten state-of-the-art COD methods on it and three general benchmarks.","keywords":["underwater camouflaged object detection","DeepCamo dataset","camouflaged object detection","SAM2 encoder","multi-scale feature enhancement","localization guidance","marine ecology monitoring"],"falsifier":"Fine-tune each of the ten baselines on DeepCamo-train under the same schedule, image size, and loss setup, then re-run Table 2 on DeepCamo-Test; if one or more baselines match or exceed SLENet, the central performance claim collapses. Independently, verify that no DeepCamo-full image overlaps any COD10K training image, since the benchmark is explicitly defined by that exclusion.","tokens_in":10148,"feed_emoji":"🐙","tokens_out":7007,"duration_ms":66457,"temperature":0.7,"pith_summary":"Underwater camouflaged object detection—finding creatures that blend into the seafloor, reefs, or water column—has no dedicated testbed, and general camouflaged-object detectors stumble on small, multiple, blurred, or color-shifted marine subjects. This paper argues that the missing piece is a benchmark plus a network that couples global localization with fine detail. It introduces DeepCamo, 2,493 underwater images of 16 species with an 8:2 train/test split and a leakage-filtered full benchmark, and reports that ten established COD methods all lose 15% or more on key metrics when moved to this data. SLENet, built on a frozen SAM2 encoder with lightweight adapters, adds a Gamma-Asymmetric Enhancement module, a Localization Guidance Branch that produces a global localization map, and a Multi-Scale Supervised Decoder that uses that map to modulate features. On DeepCamo and three public COD benchmarks, the paper reports SLENet posts the best or second-best numbers on nearly every metric, with its largest gains on DeepCamo and COD10K.","feed_headline":"SLENet tops ten models at spotting hidden marine life","feed_subtitle":"A 2,493-image benchmark and a SAM2-based detector that cuts the error rate on camouflaged sea creatures.","key_machinery":"The central machinery is a three-part guidance loop around a frozen SAM2 encoder. Gamma-Asymmetric Enhancement (GAE) is a cascaded module where each branch combines 1x1 convolutions, asymmetric convolutions, max-pooling, and rate-2 dilated convolutions; later branches receive earlier branches' outputs, letting the receptive field grow while high-resolution details are carried forward. The Localization Guidance Branch (LGB) fuses GAE-enhanced multi-scale features bottom-up into a low-resolution global localization map M. The Multi-Scale Supervised Decoder (MSSD) uses M to compute per-pixel scale and shift factors applied after batch normalization, then adds spatial attention and residual conn","core_discovery":"SLENet treats underwater camouflage as a failure of two abilities: preserving fine details across scales and knowing where the object is before refining its boundary. The network therefore freezes SAM2's Hiera-L encoder, inserts lightweight Adapters for parameter-efficient domain tuning, and feeds each of the four feature levels into a Gamma-Asymmetric Enhancement (GAE) module whose cascaded branches use asymmetric convolutions, max-pooling, and rate-2 dilated convolutions to widen receptive fields without discarding high-resolution cues. A Localization Guidance Branch (LGB) then fuses the enhanced features bottom-up into a coarse global localization map, which the Multi-Scale Supervised Dec","pith_inferences":["The paper's headline margin on DeepCamo partly reflects an asymmetric comparison: baselines are evaluated with pretrained weights while SLENet is trained on DeepCamo-train. Fine-tuning the baselines on the same split is a direct, untested way to check whether the benchmark claim is robust.","DeepCamo's combination of camouflage difficulty with optical distortion could be reused to isolate the effect of image enhancement or restoration modules: the same detector can be run on raw and enhanced versions of the same images.","Because SLENet's largest gains are on DeepCamo and COD10K, the mechanism may be more about robust multi-scale decoding under degraded inputs than about underwater-specific priors; testing on artificially degraded general COD images would separate these factors.","The localization-map-with-decay design suggests that a standalone lightweight localization branch could be plugged into other frozen foundation-model encoders as a parameter-efficient way to improve small-object segmentation."],"forward_implications":["DeepCamo can serve as a standard test for underwater camouflaged object detection, with the Table 1 numbers quantifying how far general COD models are from handling underwater scenes.","SLENet's recipe—frozen SAM2 encoder, lightweight adapters, a localization map, and a multi-scale decoder—can be applied to other fine-grained segmentation tasks where small, multiple, or boundary-complex objects are frequently missed.","The linear decay of localization-map supervision implies a training curriculum in which global location is learned early and fine details are refined later; this schedule could transfer to other guidance-based decoders.","If the reported margins survive a protocol in which baselines are fine-tuned on DeepCamo-train, SLENet would be a strong general-purpose COD method, not just an underwater specialist."],"supporting_citations":[{"why":"Supplies the SAM2 Hiera-L encoder backbone that SLENet freezes and adapts with lightweight Adapters.","marker":"[10]"},{"why":"SAM2-UNet is the closest baseline and the main runner-up; its architecture and published numbers anchor the strongest comparison in Table 2.","marker":"[8]"},{"why":"Provides the COD10K dataset and the SINet baseline, defining the general COD task and evaluation protocol that SLENet extends.","marker":"[1]"},{"why":"Supplies the CAMO dataset, used both as a source of underwater-relevant scenes for DeepCamo and as a public benchmark in Table 2.","marker":"[11]"},{"why":"Supplies the CHAMELEON dataset, used as a source of camouflaged animal images and as a public benchmark in Table 2.","marker":"[12]"},{"why":"MAS3K is one of the underwater marine-image sources from which DeepCamo's camouflaged images are curated.","marker":"[21]"},{"why":"RMAS is another marine-image source used in building the DeepCamo benchmark.","marker":"[22]"},{"why":"UFO120 contributes underwater scenes to DeepCamo, adding diverse illumination and distortion conditions.","marker":"[23]"},{"why":"HitNet is one of the ten compared baselines, representing the high-resolution iterative feedback family of COD methods.","marker":"[3]"},{"why":"PUENet is one of the ten compared baselines, representing the uncertainty-aware transformer approach to COD.","marker":"[25]"}],"fun_headline_variants":["SLENet tops ten models at finding hidden sea life","New network spots camouflaged sea creatures with guidance","DeepCamo dataset and SLENet improve underwater detection","SLENet uses multi-scale guidance to catch camouflaged marine life","Underwater camouflage no match for SLENet's enhanced vision"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The central performance claim assumes that comparing pretrained baselines against SLENet trained on DeepCamo-train is a fair protocol; if the baselines were fine-tuned on the same training split, the reported margins could shrink or reverse.","fun_headline_variants_meta":{"raw":{"variants":["SLENet tops ten models at finding hidden sea life","New network spots camouflaged sea creatures with guidance","DeepCamo dataset and SLENet improve underwater detection","SLENet uses multi-scale guidance to catch camouflaged marine life","Underwater camouflage no match for SLENet's enhanced vision"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000917,"raw_usage":{"total_tokens":3768,"prompt_tokens":736,"completion_tokens":3032,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":480,"completion_tokens_details":{"reasoning_tokens":2948}},"tokens_in":480,"tokens_out":3032,"duration_ms":23427,"temperature":1.0,"reasoning_tokens":2948,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T10:39:40.560953+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Fine-tune each of the ten baselines on DeepCamo-train under the same schedule, image size, and loss setup, then re-run Table 2 on DeepCamo-Test; if one or more baselines match or exceed SLENet, the central performance claim collapses. Independently, verify that no DeepCamo-full image overlaps any COD10K training image, since the benchmark is explicitly defined by that exclusion.","supporting_citations":[{"cited_title":"Unpublished manuscript2(6), 7 (2018)","cited_arxiv_id":null,"evidence_quote":"Supplies the CHAMELEON dataset, used as a source of camouflaged animal images and as a public benchmark in Table 2."},{"cited_title":"In: International Symposium on Benchmarking, Measuring and Optimization","cited_arxiv_id":null,"evidence_quote":"MAS3K is one of the underwater marine-image sources from which DeepCamo's camouflaged images are curated."},{"cited_title":"IEEE Journal of Oceanic Engineering (2023)","cited_arxiv_id":null,"evidence_quote":"RMAS is another marine-image source used in building the DeepCamo benchmark."},{"cited_title":"In: Pro- ceedings of the AAAI Conference on Artificial Intelligence","cited_arxiv_id":null,"evidence_quote":"HitNet is one of the ten compared baselines, representing the high-resolution iterative feedback family of COD methods."},{"cited_title":"IEEE Transactions on Image Processing 32, 3580–3591 (2023)","cited_arxiv_id":null,"evidence_quote":"PUENet is one of the ten compared baselines, representing the uncertainty-aware transformer approach to COD."}],"review_version":1}