{"id":"c2b06bee-f3cc-47d1-8231-49122bc12024","arxiv_id":"2411.18858","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"COMPrompter, a multiprompt SAM variant, reports top camouflaged object detection scores, but its main evaluation uses ground-truth prompts as input.","lead":"COMPrompter is a modified version of the Segment Anything Model that adds a second, boundary-focused prompt to detect camouflaged objects. The paper reports top scores on camouflage and polyp datasets, but those scores rely on ground-truth prompts supplied at test time, so the comparison to automatic methods is not apples-to-apples.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported SOTA relies on ground-truth-derived box/boundary prompts at inference; Table 1 compares different tasks, so the 2.2% COD10K margin is not established.","rationale":"The paper's central claim is empirical superiority. That claim requires a fair evaluation. The weakest point is not the architecture (boundary prompt as dense guidance is plausible and has internal ablation support) but the evaluation protocol: Section 3.1 explicitly uses GT at inference to construct both prompt types. This is equivalent to giving the model the ground-truth object region plus a boundary band during testing. No automatic baseline receives analogous information. The paper attempts a mitigation with 'Ours*' using UEDG-generated boundaries, but this version is not the headline result and is not SOTA. Thus the strongest claim in the abstract cannot be verified from the reported experiments. The reader's REJECT verdict is appropriate, though reframing as interactive segmentation or providing an automatic-prompt SOTA result could make the work conditionally acceptable. I agree with the reader's weakest assumption; the hyperparameter selection on test sets is an additional but secondary inflation source. Credit: the method itself is a reasonable SAM extension; EGEM/BBMG ablations show internal consistency; the precomputation and frozen encoders are sensible. But without code or weights, and with no error bars, the empirical claim is not independently checkable. The decisive test is re-evaluating with predicted prompts as primary.","tokens_in":16298,"tokens_out":4279,"duration_ms":40172,"concrete_test":"Re-run the full Table 1 comparison with the paper's own automatic-prompt pipeline only: generate the box prompt from an off-the-shelf detector (or from a baseline's predicted mask) and the boundary prompt via UEDG + binarization (the 'Our*' procedure), then compute Fωβ, Sα, Eϕ, M on all four datasets. If the automatic-prompt model does not beat MSCAF-Net and other top baselines on average, the SOTA claim in the abstract fails; the GT-prompted variant should be reported as interactive segmentation with a matching protocol.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3.1 states: 'During inference, we use the GT to generate boundaries and boxes as user prompts to simulate scenarios of user interaction.' Eq. (1) therefore injects the target mask into the boundary prompt at test time, and the box prompt is also taken from GT. The main comparison in Table 1 is against automatic COD methods (UCNet, SINet, MSCAF-Net, etc.) that receive no GT-derived location or boundary information; COMPrompter is thus solving an easier, partially supervised task. The predicted-prompt variant the paper supplies ('Ours*', built from UEDG edges plus binarization) is explicitly said to rank only 'within the top three,' not first. Since the abstract's 'cutting-edge performance' and the 2.2% COD10K margin refer to the GT-prompted model, the central SOTA claim is unsupported. Additional issues compound this: Table 5 chooses dilation kernels and Table 6 chooses the binarization offset by comparing metrics on the same test sets, and no error bars or code/weights are provided. The internal ablations (M1→M3, M4→M5) do show the boundary prompt and DWT help, but those gains do not rescue the unfair comparison.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes COMPrompter, a SAM-based multiprompt network for camouflaged object detection. It introduces a boundary prompt computed by an edge gradient extraction module (EGEM), a box-boundary mutual guidance module (BBMG), and discrete wavelet transform high-frequency features. The model is evaluated on CAMO, CHAMELEON, COD10K, NC4K, and five polyp segmentation datasets, with claims of state-of-the-art performance. The main quantitative results use ground-truth-derived boundary and box prompts at inference; a separate practical variant (Ours*) uses automatically generated boundaries and is reported to rank within the top three.","tokens_in":16470,"tokens_out":5784,"duration_ms":49317,"significance":"If the method were shown to outperform prior automatic methods under identical task conditions, the multiprompt design and frequency-domain augmentation would be of interest to COD and SAM-adaptation research. The internal ablations indicate that each proposed component contributes in the prompted setting, and the precomputation of image embeddings is an efficiency measure worth noting. However, the empirical case for state-of-the-art performance is not established because the comparison conditions differ between the proposed method and the baselines, and because key hyperparameters are tuned on the test benchmarks.","major_comments":[{"comment":"The main comparison in Table 1 does not measure the same task for COMPrompter and the baseline methods. Section 3.1 states that during inference \"we use the GT to generate boundaries and boxes as user prompts to simulate scenarios of user interaction,\" and Eq. (1) defines the boundary prompt as BG = gamma(mu(GT) - GT) * C(I), so the target mask is injected into the model at test time. The automatic COD methods listed in Table 1 receive no such target-derived location or boundary information, and the SAM, SAM-Adapter, and MedSAM rows do not appear to receive the same box and boundary prompts. Consequently, the claimed SOTA margins, including the abstract's 2.2% average positive gain on COD10K, are not supported. The paper's own automatic variant (Ours*, built from UEDG edges plus binarization) is described in Section 4.3 as ranking only \"within the top three,\" which is consistent with this concern. A fair comparison would need either GT-prompted baselines (e.g., SAM and MedSAM given GT boxes and boundaries) or a fully automatic prompt source for all methods.","section":"Section 3.1, Eq. (1); Table 1"},{"comment":"The dilation kernel sizes (D1-D5) in Table 5 and the binarization offset value in Table 6 are selected by comparing metrics on the same test datasets (COD10K, NC4K, CAMO, CHAMELEON) used for the final claim. This post hoc selection on the test benchmarks means the reported results are partly fitted; the paper does not provide a held-out validation split, error bars, or repeated-run statistics. The ablation in Table 3 therefore demonstrates internal usefulness of the proposed modules only under this test-set-selected configuration and does not quantify the stability or generalization of the reported gains.","section":"Tables 5 and 6; Section 4.4"},{"comment":"The polyp segmentation comparison in Table 2 is subject to the same GT-prompting issue as the COD comparison. The global inference protocol of Section 3.1 (\"we use the GT to generate boundaries and boxes as user prompts\") appears to apply to the polyp experiments as well, so the paper does not establish that COMPrompter outperforms the listed polyp methods under an equivalent task definition. The internal ablations (M1 to M5 in Table 3) are informative about the relative contribution of the box prompt, boundary prompt, and DWT once GT prompts are available, but they do not justify the headline SOTA claims.","section":"Table 2; Section 3.1"}],"minor_comments":[{"comment":"The section title \"Datatset\" is a typo for \"Dataset.\"","section":"Section 4.1"},{"comment":"The header \"CVC-linicDB\" is a typo for \"CVC-ClinicDB,\" which is the name used elsewhere in the paper.","section":"Table 2"},{"comment":"The scatter plot in Figure 1 is difficult to read because the x-axis is MAE and the y-axis is the sum of three metrics with different scales; the caption should state any normalization explicitly, and Figure 1 also contains the typo \"T est.\"","section":"Figure 1"},{"comment":"There are several grammatical slips, including \"achieves a the cutting-edge performance\" in the contributions and \"It use a multiprompt strategy\" in the Conclusion.","section":"Abstract and Conclusion"},{"comment":"The \"Ours*\" column is introduced in Table 1 without a table footnote; the explanation of the star appears only later in Section 4.3, which makes the table initially confusing.","section":"Table 1"},{"comment":"The abstract promises code availability, but no code or pretrained weights are provided at this time; please clarify the release status, as this hampers reproducibility of the reported numbers.","section":"Code availability"}],"recommendation":"reject","confidential_remarks":"The manuscript's central claim depends on an evaluation protocol in which the proposed method receives ground-truth-derived prompts while automatic baselines do not. This is not a minor presentational flaw: the headline numbers in the abstract and Table 1 are not supported by the reported experiments. The test-set-based selection of dilation and offset parameters and the absence of error bars further weaken the evidence. I would encourage reconsideration if the authors can provide a fully automatic evaluation with a held-out validation split, GT-prompted baselines, and released code; as currently written, the contribution is not publishable in its present form."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The COMPrompter paper is a good example of a plausible architectural idea undermined by an unfair evaluation. The multiprompt design — box plus a gradient boundary prompt, fused via BBMG with DWT high-frequency features — is genuinely new, and the ablations show each component helps. The M1→M3 jump from the boundary prompt is large, and the DWT contribution, while modest, is consistent. I believe the architecture is a reasonable extension of SAM for COD.\n\nThe problem is the main comparison. Section 3.1 states they use GT to generate boundaries and boxes as user prompts at inference. That gives COMPrompter access to the target mask, while all automatic baselines get nothing. Table 1 thus compares different tasks: COMPrompter is doing interactive-style segmentation with a perfect box and boundary, while the baselines are doing fully automatic COD. The 2.2% COD10K margin is therefore not evidence of a better COD model. The paper does include a predicted-prompt variant (Ours*) using UEDG plus binarization, and that version ranks only within the top three, not first. That discrepancy is telling.\n\nThe hyperparameter selection is also post hoc: dilation kernels in Table 5 and offset in Table 6 are chosen by testing on the same benchmarks, and no error bars or code/weights are available. That compounds the uncertainty. The conclusion's own admission about multiple objects in one box is honest but adds another limitation.\n\nIf the authors reframe this as an interactive segmentation method, or evaluate with predicted prompts (from a detector or UEDG) and report error bars, the central claim could be restored. The internal ablations already suggest the boundary prompt helps when the prompt is given; the open question is whether it helps when the prompt is imperfect. As submitted, the SOTA claim fails, but the idea deserves a serious referee rather than a desk reject.","headline":"A clever SAM multiprompt design, but the headline SOTA margin relies on GT-derived prompts at inference and does not survive contact with an automatic comparison.","tokens_in":17049,"tokens_out":1405,"would_cite":false,"duration_ms":12675,"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":"Multiprompt SAM variant reaches state-of-the-art camouflaged object detection","keywords":["camouflaged object detection","segment anything model","multiprompt network","boundary prompt","edge gradient extraction","box-boundary mutual guidance","discrete wavelet transform","polyp segmentation"],"falsifier":"Run COMPrompter on COD10K with box and boundary prompts generated automatically (for example, from a pretrained detector and the paper's UEDG pipeline with offset 15) instead of from ground truth, and compare it against the same competitors under identical conditions; if the reported margin over the next-best method shrinks below statistical significance or inverts, the state-of-the-art claim does not survive real deployment.","tokens_in":15993,"feed_emoji":"🔍","tokens_out":2720,"duration_ms":26431,"temperature":0.7,"pith_summary":"This paper tries to show that the Segment Anything Model can be turned into a leading camouflaged object detector by replacing its single box prompt with a multiprompt strategy. The new network, COMPrompter, adds a boundary prompt that encodes edge-gradient information and fuses it with the original box prompt, plus high-frequency features from a discrete wavelet transform. If correct, the result is the best published scores on four COD benchmarks and five polyp segmentation datasets, including an average 2.2% improvement over the prior leading model on COD10K. The claim matters because it points a way to adapt large segmentation foundation models to difficult, fine-structured tasks without retraining their heavy image encoders.","feed_headline":"Boundary-prompt SAM tops camouflaged object detection","feed_subtitle":"Adding gradient edge prompts to box prompts pushes scores past prior leaders on four COD and five polyp benchmarks.","key_machinery":"The load-bearing mechanism is the boundary prompt generated by EGEM, following $BG = \\gamma(\\mu(GT) - GT) * C(I)$, where $\\mu$ is a $3\\times3$ dilation, $\\gamma$ a $5\\times5$ dilation, and $C$ a Canny operation. This gradient-bearing boundary mask is encoded by a frozen prompt encoder and combined with the dense box embedding through BBMG, which uses adapted pointwise depthwise-separable convolution with residual connections; the diagonal high-frequency component $HF = x_1 - x_2 - x_3 + x_4$ from the discrete wavelet transform is then concatenated and convolved to form the final dense embedding that guides the mask decoder.","core_discovery":"COMPrompter establishes that a SAM-based detector can surpass specialized camouflaged object detection methods when the prompt is enriched rather than left as a single box. The paper proposes a boundary prompt built by an edge gradient extraction module that dilates the ground-truth mask, subtracts the original mask, and multiplies the result with a Canny edge map of the image, producing a boundary mask that carries gradient information at the object-background junction. A box-boundary mutual guidance module then fuses dense box and boundary embeddings, and diagonal high-frequency features from a discrete wavelet transform are appended as a supplementary signal. With the image encoder and prompt encoder frozen and only the mask decoder fine-tuned, the method reports state-of-the-art results on CAMO, CHAMELEON, COD10K, and NC4K, and on five polyp segmentation benchmarks, while cutting SAM's parameter count to 94.86M and quadrupling its inference speed.","pith_inferences":["The reported superiority depends on ground-truth-derived prompts at inference; a practical deployment would need an automatic box and boundary generator, so the approach is best read as a prompt-design result rather than a fully autonomous detector.","The boundary-prompt mechanism is domain-agnostic and could plausibly transfer to other edge-critical segmentation tasks such as aerial imagery, satellite building extraction, or lesion boundary delineation, provided the dilation kernels and binarization offset are retuned.","The paper's own note that a single box can miss multiple targets suggests a natural extension: using several sub-boxes with per-box boundary prompts, which would also make the method more robust to multi-instance scenes.","If a paired test were run where competitors also receive the same box and boundary prompts, the comparison would isolate the value of the multiprompt fusion itself; current tables conflate prompt advantage with network advantage."],"forward_implications":["Any SAM-based camouflaged object detector could adopt the boundary prompt and expect a large jump: adding the boundary prompt to box-only SAM raises weighted F-measure by 5.8 percentage points on average across COD10K and NC4K.","The diagonal high-frequency (HH) component of the DWT is the useful band for COD; it outperforms LL, LH, and HL on all four metrics, so future SAM adaptations can focus on that component.","The approach transfers to polyp segmentation, where it reports gains of about 3.2% in mean Dice and 4.5% in mean IoU over MedSAM across the five evaluated datasets.","Because the heavy encoders are frozen and precomputed, the method is trainable in about 4.2 hours on a single RTX 3080Ti, suggesting foundation-model fine-tuning for COD is feasible without large compute budgets."],"supporting_citations":[{"why":"Supplies the base Segment Anything Model that COMPrompter modifies, including its frozen image encoder, prompt encoder, and mask decoder.","marker":"[21]"},{"why":"Provides the precedent of fine-tuning only the mask decoder with a box prompt in medical images, which COMPrompter extends to COD.","marker":"[20]"},{"why":"Defines the COD10K dataset, the training/test split, and the four evaluation metrics that anchor the main quantitative claims.","marker":"[13]"},{"why":"Inspires the use of discrete wavelet transform for feature decomposition in camouflaged object detection, justifying the high-frequency supplement.","marker":"[23]"},{"why":"Is the SAM-Adapter baseline, a SAM-based COD method that COMPrompter must beat to justify its added complexity.","marker":"[19]"},{"why":"Provides the UEDG edge detector used to generate automatic boundaries for the Our* variant, which is the closest the paper comes to prompt-free evaluation.","marker":"[36]"},{"why":"Supplies the depthwise separable convolution used in the box-boundary mutual guidance module.","marker":"[22]"}],"fun_headline_variants":["Expanded SAM prompts eclipse camouflaged object detectors","Gradient edge prompts lift SAM beyond camouflage baselines","Multiprompt SAM with wavelet edges tops COD charts","Boundary and box prompts push SAM to COD SOTA","SAM with dual prompts beats leading COD models by 2.2%"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The benchmark scores assume that ground-truth-derived box and boundary prompts are a fair stand-in for real user prompts, while all automatic comparison methods must detect the camouflaged object without that target-specific information.","fun_headline_variants_meta":{"raw":{"variants":["Expanded SAM prompts eclipse camouflaged object detectors","Gradient edge prompts lift SAM beyond camouflage baselines","Multiprompt SAM with wavelet edges tops COD charts","Boundary and box prompts push SAM to COD SOTA","SAM with dual prompts beats leading COD models by 2.2%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000642,"raw_usage":{"total_tokens":2986,"prompt_tokens":1010,"completion_tokens":1976,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":626,"completion_tokens_details":{"reasoning_tokens":1895}},"tokens_in":626,"tokens_out":1976,"duration_ms":12204,"temperature":1.0,"reasoning_tokens":1895,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T10:48:55.478501+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run COMPrompter on COD10K with box and boundary prompts generated automatically (for example, from a pretrained detector and the paper's UEDG pipeline with offset 15) instead of from ground truth, and compare it against the same competitors under identical conditions; if the reported margin over the next-best method shrinks below statistical significance or inverts, the state-of-the-art claim does not survive real deployment.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the COD10K dataset, the training/test split, and the four evaluation metrics that anchor the main quantitative claims."},{"cited_title":"Sifre, ``Rigid-motion scattering for image classification,'' Ph.D","cited_arxiv_id":null,"evidence_quote":"Inspires the use of discrete wavelet transform for feature decomposition in camouflaged object detection, justifying the high-frequency supplement."},{"cited_title":"Zhang, D.-P","cited_arxiv_id":null,"evidence_quote":"Is the SAM-Adapter baseline, a SAM-based COD method that COMPrompter must beat to justify its added complexity."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the UEDG edge detector used to generate automatic boundaries for the Our* variant, which is the closest the paper comes to prompt-free evaluation."},{"cited_title":"Kirillov, E","cited_arxiv_id":null,"evidence_quote":"Supplies the depthwise separable convolution used in the box-boundary mutual guidance module."}],"review_version":1}