{"id":"90ae5a31-8366-403d-a4eb-4c1e11517034","arxiv_id":"2505.09123","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A lightweight point-prompt generator enables frozen SAM to reach state-of-the-art camouflaged object detection on several benchmarks.","lead":"A small network picks good and bad spots on an image, then a frozen SAM foundation model segments camouflaged objects that it previously failed to find. If correct, this offers a cheap recipe for adapting large segmentation models to hard niche tasks without retraining them.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Headline Table 2 gain does not isolate selective point guidance from mask feedback or naive point prompts; a random/trivial-prompt control is required to support the central claim.","rationale":"The paper's central claim—frozen SAM can be made competitive for COD by automatic selective positive/negative point prompts—is plausible, and the quantitative jump over unprompted SAM is large. I do not object to using a lightweight trained network to propose prompts; that is a legitimate design. The load-bearing weakness is evidentiary rather than architectural: the reported gain bundles prompt generation, point selection, and SAM's two-pass mask refinement. The ablations decompose the second pass and the number of negative points but never compare the point-selection algorithm against a cheap or random prompt policy. Since SAM is explicitly designed to turn sparse points into masks, a trivial inside/outside prompt might recover most of the gain; if so, the paper's 'selective key point' contribution would be unsupported, even though the broader 'SAM can be prompted for COD' statement might survive. I also note the Fm SOTA statement is internally contradicted by Table 1 on CAMO, and hyperparameters appear to be chosen on test sets; these reinforce the need for the prompt-control experiment, but the missing control is the single most load-bearing issue. This does not invalidate the paper; it makes the central claim conditional on a plausible and easily run control experiment.","tokens_in":11577,"tokens_out":5153,"duration_ms":53757,"concrete_test":"On the same COD10K-train/NC4K-test and COD10K-test protocol, run the identical two-pass SAM pipeline replacing PPT-net/KPS prompts with three controls: (i) a uniformly random 16×16 grid of one positive and two adjacent negatives, (ii) one positive at the centroid of the largest high-probability cluster from PPT-net plus its two nearest L4 negatives, skipping farthest-point PP selection, and (iii) oracle positives/negatives sampled from the ground-truth mask. Report S-measure, Fm, and M for each. If control (ii) or even (i) stays within 0.01 S-measure of the full method, the claimed selectivity of KPS is not load-bearing; if the full method exceeds controls by a margin that survives held-out hyperparameter selection, the central claim is supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Table 2 reports Sm 0.406→0.836 on NC4K and 0.445→0.826 on COD10K for 'SAM vs SAM+prompt'. However, Section 3.4 defines SAM+prompt as two SAM passes: point triples, then the first mask fed back as an additional mask prompt (SAMAug-style). Table 5 shows the second pass contributes only ~0.005 Sm, and Table 4 shows 10 NPs add ~0.003, so the remaining ~0.41 gain is attributed implicitly to the prompt points. No experiment compares PPT-net/KPS prompts to a trivial control: e.g., uniformly sampled positive/negative points, one positive at the centroid of the highest-score cluster, or ground-truth-derived oracle points. Because SAM is a strong promptable segmenter and the first mask is available to the second pass, the observed improvement could in principle be obtained by SAM's own segmentation plus almost any point inside/outside the object, making the 'selective key point' contribution unverified. The central claim—automatic selective key point guidance enables SAM for COD—therefore rests on an uncashed assumption that the prompt generation, not the mask-feedback mechanism or SAM's internal generalization, is responsible for the result. This is compounded by test-set-based choice of K, grid resolution, and iteration count; a held-out validation split is not reported.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper addresses the question of whether the Segment Anything Model (SAM) can be made to work for Camouflaged Object Detection (COD) through automatically generated point prompts. It proposes a two-part prompt generation pipeline: PPT-net, a lightweight network that predicts, for each cell of a uniform M×M grid, the probability that a camouflaged object is present at the cell center; and a Key Point Selection (KPS) algorithm that converts these probability scores into triples of one positive and two negative points. The prompts are fed to a frozen SAM, followed by a second SAM pass in which the first mask is used as a mask prompt, following the SAMAug idea. Experiments on COD10K, NC4K, and CAMO compare the method against specialized COD baselines and report large gains over unprompted SAM (e.g., NC4K S-measure from 0.406 to 0.836, Table 2). The paper claims new state-of-the-art Fm_beta on all three datasets and Fw_beta on two datasets, and positions the work as the first to facilitate a big model for COD.","tokens_in":11903,"tokens_out":8864,"duration_ms":89626,"significance":"If the central attribution is verified, the paper makes a useful practical point: a frozen generalist segmenter, guided by inexpensive point-level localization, can approach specialized COD models, reducing the need for heavily engineered COD architectures. The quantitative effect in Table 2 is striking, and the training setup is honest in an important respect: PPT-net is trained on averaged ground-truth masks of the training split, SAM is frozen, and the final mask comes from SAM on held-out test images, so there is no self-referential circularity. The main gap is that the experiments do not isolate the proposed selective key-point mechanism from trivial prompt baselines; the missing controls are essential before the paper can claim that selective key points, rather than SAM's own promptable segmentation capability, are responsible for the observed improvement. The SOTA claims also need to be reconciled with the paper's own Table 1.","major_comments":[{"comment":"The paper states in §4.2 that \"our approach establishes a new state-of-the-art in terms of Fm_beta on all the 3 datasets,\" but Table 1 does not support this. On CAMO, the proposed method obtains Fm_beta=0.790 while ZoomNet obtains 0.794; on the same dataset Fw_beta is 0.742, tied with SiNetV2 and below ZoomNet's 0.752. The abstract's claim of results \"over the existing methods on 3 data sets under 6 metrics\" is also broader than what the table shows. Please correct the SOTA claims to name the exact metrics and datasets on which the method leads, and discuss the CAMO shortfall.","section":"§4.2, Table 1"},{"comment":"The headline evidence (Table 2: NC4K S-measure 0.406 to 0.836) compares unprompted SAM with the full pipeline, which includes both the KPS point triples and the second SAM pass with mask feedback. The ablations show that the second pass adds only about 0.005 in S-measure (Table 5) and the 10 negative points add about 0.003 (Table 4), but no experiment isolates the proposed PPT-net/KPS prompt generator from trivial prompt baselines. A control with uniformly sampled grid points, a single centroid positive point, or oracle points derived from the ground-truth mask is needed to verify that \"selective key point\" selection, rather than SAM's own ability to segment from any interior point, is responsible for the gain. Without such controls, the central attribution in the title and contributions is not established.","section":"§3.4, Table 2 vs. Tables 4/5"},{"comment":"Algorithm 1 is not executable as written because line 4 computes arg max over d(C_i, P) while P is initialized to the empty set; the distance to an empty set is undefined, so the first positive point cannot be selected by the stated rule. The surrounding text says the first PP is the point with the highest prediction score, but this rule does not appear in the pseudocode. Also, lines 17-18 use d(C_i, P_j) with P_j a single point, while the earlier definition d(C_i, P) is for a collection; please clarify the notation for point-to-point and point-to-set distances. These issues make the central selection algorithm non-reproducible as described.","section":"§3.3, Algorithm 1"},{"comment":"The ablations select the grid size, number of positive/negative points, and number of SAM iterations by performance on the COD10K test split, but the paper does not describe a held-out validation split or a hyperparameter selection procedure. For example, Table 3 shows that 32×32 candidates outperform 16×16 (S-measure 0.834 vs. 0.826), yet the main results use 16×16. Since these hyperparameters directly determine the prompts, the reported numbers may reflect test-set tuning; please report how K, M, the confidence thresholds in Eq. (6), and the number of iterations were chosen, and provide sensitivity results on NC4K and CAMO.","section":"§4.3, Tables 3-5"}],"minor_comments":[{"comment":"The term \"promotion\" is used throughout where \"prompt\" is standard (e.g., abstract, §1, Fig. 2 caption); please define this terminology at first use to avoid confusion.","section":"Throughout"},{"comment":"There is a typo in \"receptive filed\" which should be \"receptive field,\" and in §4.3 \"has a higher chance\" should be \"have a higher chance.\"","section":"§3.3"},{"comment":"References [19] and [20] are the same paper (Le et al., Anabranch network for camouflaged object segmentation); please merge them.","section":"References"},{"comment":"Please specify the exact train/test split of COD10K (the table reports 2,026 test images) and clarify how NC4K and CAMO are used for evaluation; the sentence \"we train our method only on COD10K data set\" is ambiguous.","section":"§4.1"},{"comment":"The claim that this is \"the first work to facilitate big model for COD\" is not substantiated against prior prompt-based SAM-COD adaptations; please cite and discuss the relevant literature or soften the novelty claim.","section":"Abstract and §1"}],"recommendation":"major_revision","confidential_remarks":"To the editor: the core quantitative finding in Table 2 is likely real and worth pursuing, but the central attribution to selective key-point guidance is not yet isolated, and the state-of-the-art claims are internally inconsistent with Table 1. I recommend major revision rather than rejection because the missing control experiments and corrected claims are straightforward to address within the manuscript's scope."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know before you spend time on this. The load-bearing result is Table 2: unprompted SAM gets 0.406 S-measure on NC4K, and SAM with the proposed point prompts gets 0.836. That is a striking gain and it is probably real, because SAM is frozen and the result comes from the prompt pipeline. The second thing is that the paper has not shown that its specific key point selection is responsible for the gain. There is no control with random points, centroid points, or oracle points. Given how strong SAM is as a promptable segmenter, the improvement could in principle come from almost any point inside the object plus the contrastive negatives. The selective KPS contribution is unverified.\n\nWhat is genuinely new: the PPT-net + KPS pipeline as a lightweight front-end for frozen SAM, and the empirical reversal of the earlier claim that SAM is unsuitable for COD. The authors deserve credit for keeping SAM frozen and for showing that a simple probability grid can replace a bespoke COD network. The mask feedback worry from the stress-test does not land: Table 5 shows the second SAM pass adds only about 0.005 S-measure, so the gain is almost entirely from the point prompts themselves.\n\nThe soft spots are real but uneven. The SOTA overstatement is concrete: the text says new SOTA in Fm_beta on all three datasets, but Table 1 shows ZoomNet at 0.794 vs. our 0.790 on CAMO, and Em/Ex are below the best baselines everywhere. That should be fixed. The hyperparameters (grid size, number of points, thresholds, iterations) appear to be tuned on the test set without a reported validation split, and there are no error bars or code release. Algorithm 1 has a small but real gap: the first arg max calls d(C_i, P) when P is empty. None of these are fatal, and the mask-feedback concern is not, but together they mean the central claim is conditional.\n\nWho should read it: anyone working on SAM prompting or on COD who wants a concrete example of how to redirect a frozen foundation model. It deserves a serious referee, not a desk reject, because the core result is plausible and the method is simple enough to test. The revision needs a trivial-prompt ablation, a validation split, corrected SOTA claims, and code. Without those, I would not cite it as evidence that selective key point guidance works.","headline":"The headline jump in Table 2 looks real, but the paper does not prove that its selective key point generation is what causes it, and one SOTA claim is contradicted by its own table.","tokens_in":12414,"tokens_out":3033,"would_cite":false,"duration_ms":33850,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A frozen SAM, prompted with automatically selected contrastive points, matches specialized camouflaged-object detectors, raising NC4K S-measure from 0.406 to 0.836.","keywords":["camouflaged object detection","Segment Anything Model","visual prompting","point prompts","key point selection","multi-scale feature prediction","prompt-based segmentation","frozen foundation model"],"falsifier":"Rerun the Table 2 experiment with the selected points replaced by randomly sampled points of the same count inside and outside the object: if random prompts recover most of the 0.406-to-0.836 S-measure gain on NC4K, the key-point selection algorithm is not the driver; if they recover almost none, SAM is genuinely relying on the contrastive triples.","tokens_in":11342,"feed_emoji":"🎯","tokens_out":11386,"duration_ms":102662,"temperature":0.7,"pith_summary":"Camouflaged object detection, the task of finding objects that blend into their surroundings, was previously judged too hard for the Segment Anything Model (SAM). This paper sets out to overturn that conclusion, arguing that SAM fails only because it is not prompted properly. The authors build a lightweight prompt generator, PPT-net plus a key-point selection algorithm, that places a positive point on the object and two nearby negative points on the background, and they feed those triples to a frozen SAM. In their experiments, this raises SAM's S-measure on the NC4K test set from 0.406 to 0.836 and yields results comparable to, and on some precision/recall metrics better than, specialist camouflaged-object detectors. If correct, the result matters because COD becomes a coarse localization problem rather than a pixel-perfect boundary problem, letting a frozen general-purpose segmenter replace task-specific architectures.","feed_headline":"Point prompts turn Segment Anything into a camouflaged-object detector","feed_subtitle":"A light selector picks positive and negative points, lifting NC4K S-measure from 0.406 to 0.836.","key_machinery":"The load-bearing mechanism is the contrastive point-prompt triple produced by two cooperating components. PPT-net (Promotion Point Targeting Network) predicts, for each cell of an M-by-M grid, the probability that a camouflaged object occupies that cell, using multi-scale features from a Res2Net encoder augmented by receptive-field blocks. KPS (Key Point Selection) then chooses high-confidence, mutually distant cells as positive points, and for each positive point selects the two nearest low-confidence cells as negative points, bracketing the object's edge. Feeding these triples to SAM, a general-purpose segmentation network whose weights are never updated, converts COD from exact segmentation into coarse localization; a second SAM pass with the first mask as an additional prompt repairs boundary detail.","core_discovery":"The paper's central claim is that SAM's poor performance on camouflaged objects is a prompting failure, not an intrinsic limit. The mechanism is a triple of point prompts per selected location: one positive point sampled from high-confidence predictions, and two negative points from the nearest low-confidence grid cells. These triples tell SAM where the object is and roughly where its edge lies, and a second SAM pass uses the first output mask as an additional prompt to refine boundaries. On the paper's reported numbers, NC4K S-measure rises from 0.406 to 0.836, and the method reaches the top of the baseline table on the weighted and mean F-measures for the large datasets while remaining below the best edge-focused E-measures, which the authors attribute to the finite number of prompt points covering only part of the boundary.","pith_inferences":["A natural experiment the paper does not run is a sensitivity check: if selected key points are perturbed by one grid cell, a large S-measure drop would show the precision of the point selector matters more than SAM's tolerance for coarse prompts.","The same contrastive-point recipe may transfer to other low-contrast segmentation tasks, such as medical lesions or objects in satellite imagery, where a coarse locator could replace a domain-specific decoder.","The grid-density comparison suggests an adaptive or non-uniform candidate grid concentrated near predicted object locations could cover small or fragmented objects better than a fixed uniform grid.","The authors attribute some remaining errors to objects split into isolated parts, which points to a specific testable weakness: recall on multi-part objects may drop because KPS spreads only a handful of positive points across the entire image."],"forward_implications":["If the claim holds, COD no longer requires hand-crafted edge-, wavelet-, or frequency-domain modules: a frozen general-purpose segmenter plus a small point selector reaches comparable performance.","SAM's documented struggles in concealed scenes become a prompt-design problem rather than evidence that the model cannot handle the task.","On the reported metrics, the largest gains are in precision/recall-weighted measures such as weighted and mean F-measure, so the method's strength is capturing overall shape rather than exact edges.","Because SAM stays frozen, adapting to a new concealment setting only requires training the lightweight PPT-net on grid-level probabilities, which is cheaper than training a specialized COD model from scratch."],"supporting_citations":[{"why":"Supplies the frozen SAM model and its point-prompt interface; all segmentation output comes from SAM.","marker":"[18]"},{"why":"Earlier study declaring SAM ineffective in concealed scenes; its conclusion is the claim this paper overturns.","marker":"[15]"},{"why":"Second earlier study reporting SAM failure on camouflaged object detection; the target of the reversal.","marker":"[34]"},{"why":"Defines the COD task and provides the COD10K training/test set and the SiNet baseline.","marker":"[9]"},{"why":"Provides the NC4K test set on which the largest S-measure jump, from 0.406 to 0.836, is measured.","marker":"[25]"},{"why":"Supplies the Res2Net encoder weights used to initialize PPT-net and serves as a baseline.","marker":"[8]"},{"why":"Provides the SAMAug iterative prompt-feedback procedure reused for the second SAM pass with the mask prompt.","marker":"[5]"},{"why":"FEDER, a recent COD baseline that the reported method beats on weighted and mean F-measure on the large datasets.","marker":"[11]"}],"fun_headline_variants":["Point prompts make SAM work for camouflaged objects","Triple-point prompts turn SAM into a camouflage finder","Key-point guidance lifts SAM's S-measure to 0.836","Selective prompts unlock SAM for camouflaged detection"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The entire pipeline depends on a handful of coarse positive and negative points carrying enough boundary information for SAM to produce a precise, full-boundary mask of a camouflaged object; if point prompts cannot do that for camouflage, the frozen SAM has no way to recover and the method collapses.","fun_headline_variants_meta":{"raw":{"variants":["Point prompts make SAM work for camouflaged objects","Triple-point prompts turn SAM into a camouflage finder","Key-point guidance lifts SAM's S-measure to 0.836","Selective prompts unlock SAM for camouflaged detection"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000264,"raw_usage":{"total_tokens":1594,"prompt_tokens":924,"completion_tokens":670,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":540,"completion_tokens_details":{"reasoning_tokens":602}},"tokens_in":540,"tokens_out":670,"duration_ms":7002,"temperature":1.0,"reasoning_tokens":602,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T21:38:56.687665+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Rerun the Table 2 experiment with the selected points replaced by randomly sampled points of the same count inside and outside the object: if random prompts recover most of the 0.406-to-0.836 S-measure gain on NC4K, the key-point selection algorithm is not the driver; if they recover almost none, SAM is genuinely relying on the contrastive triples.","supporting_citations":[{"cited_title":"In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition","cited_arxiv_id":null,"evidence_quote":"Defines the COD task and provides the COD10K training/test set and the SiNet baseline."},{"cited_title":"In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition","cited_arxiv_id":null,"evidence_quote":"Provides the NC4K test set on which the largest S-measure jump, from 0.406 to 0.836, is measured."},{"cited_title":"IEEE transactions on pattern analysis and machine intelligence44(10), 6024–6042 (2021) 4, 8, 10, 11","cited_arxiv_id":null,"evidence_quote":"Supplies the Res2Net encoder weights used to initialize PPT-net and serves as a baseline."},{"cited_title":"In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition","cited_arxiv_id":null,"evidence_quote":"FEDER, a recent COD baseline that the reported method beats on weighted and mean F-measure on the large datasets."}],"review_version":1}