{"id":"6c5b1da3-dd32-46b3-b43a-886a586fd059","arxiv_id":"2607.27843","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A depth-collaborative network using prototype contrastive learning reports state-of-the-art camouflaged object detection on CAMO, COD10K, and NC4K.","lead":"A new neural network combines color images and depth maps, using learned 'prototype' tokens to keep the two modalities distinct while fusing them, and reports top scores for camouflaged object detection. The paper's claimed gains are plausible but one table entry looks like a possible typo and no error bars are reported.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"COD10K Fwβ = 0.930/0.932 in Table 1 is internally implausible and unsupported by other metrics; if this is a typo or column misalignment, the SOTA claim loses its main evidence.","rationale":"The reader's weakest_assumption is pseudo-depth quality, but I think the Fwβ anomaly is more load-bearing because the SOTA claim is decided by Table 1 and a single metric jump can be checked directly. Pseudo-depth reliability affects interpretation/transferability but not the fairness of same-depth comparisons; even if DAV2 pseudo-depth is imperfect, VCP-DCN is compared against Samba/DaCOD using the same depth source, so the relative SOTA claim survives unless depth noise interacts differently with architectures. The COD10K Fwβ outlier, by contrast, is an internal inconsistency: the same model gets 0.871 Fw on CAMO/NC4K but 0.930 on COD10K despite lower Sm. This looks like a transposed digit or misaligned column. The proposed test (recompute that one metric from public artifacts) would settle it. I therefore keep the reader's CONDITIONAL verdict: the architecture/ablation story is coherent, but the headline SOTA number needs verification before acceptance.","tokens_in":15694,"tokens_out":7061,"duration_ms":64551,"concrete_test":"Recompute the COD10K Fwβ for both VCP-DCN rows using the released code/checkpoints and the standard CODToolbox metric, without changing any preprocessing. Compare each recomputed value to 0.930 and 0.932. If the recomputed value is below 0.90 or differs by more than 0.01, the reported SOTA claim is not supported as stated. If the code/checkpoints are not sufficient to reproduce the table, request the authors' raw prediction maps for COD10K and evaluate those; either way, this single number determines whether the headline comparison stands.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim — 'VCP-DCN outperforms previous SOTA COD methods' — is carried by Table 1. In that table, VCP-DCN reports COD10K Fwβ = 0.930 (VMamba) and 0.932 (Swin) when pseudo-depth comes from Depth Anything V2. The best same-backbone comparison, Samba with the same depth source, reports 0.842; the next-best RGB-D methods are near 0.80. No other row in the paper shows a comparable jump: on CAMO and NC4K, VCP-DCN's Fwβ is 0.871 with Sm≈0.89–0.90, while on COD10K it reports Fwβ 0.930 with Sm≈0.88. A +0.088 Fwβ improvement, accompanied by only a 0.002 MAE change and no corroborating jump in Sm or Eϕ on COD10K, is far outside the normal COD spread and not explained by any component ablation (the largest ablation gain is ~0.03 Fwβ). If the 0.930/0.932 entries are typos or column misalignments, the 'consistently outperforms' statement loses its strongest quantitative support on the largest benchmark; if they are genuine, the paper provides no error bars, multiple seeds, or runnable evaluation config (the released repo has no commit hash or config files) to verify them. This is an internal-consistency/reproducibility concern at the exact point where the SOTA conclusion is decided.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes VCP-DCN, an RGB-D camouflaged object detection (COD) method that introduces a three-stage depth collaborative network: a Separable Prototype Embedding (SPE) module for aligning RGB and depth features through modality-consistency and modality-specific prototype contrastive learning, a Multi-modality Dual Attention (MDA) module for cross-modal feature interaction, and a Depth Adaptive Injection (DAI) module for adaptive fusion. The method is evaluated on CAMO, COD10K, and NC4K with pseudo-depth from Depth Anything V2, and the authors claim consistent state-of-the-art performance over existing RGB and RGB-D COD methods, with additional experiments on RGB-D salient object detection (SOD). The architecture is coherent and the ablations indicate that each component contributes, but the central quantitative claim is undermined by an unexplained and internally inconsistent set of Fwβ scores in Table 1.","tokens_in":16080,"tokens_out":5744,"duration_ms":49426,"significance":"If the reported results are reliable, VCP-DCN would represent a genuine advance in RGB-D COD, demonstrating the value of explicitly separating modality-consistency and modality-specific prototypes in a depth collaborative framework. The extension to RGB-D SOD also suggests that the modules are transferable. The design is well motivated and the ablation study supports the utility of the three proposed modules. However, the paper's headline conclusion hinges on Table 1, where a suspicious COD10K Fwβ value (0.930/0.932) is far outside the range of all comparable methods and is not corroborated by other metrics. The lack of reproducibility details (no commit hash, no evaluation config, no multiple seeds) further weakens confidence. The central claim is therefore currently unverified, and the manuscript requires substantial revision to either justify or correct the anomalous numbers.","major_comments":[{"comment":"In the 'RGB-D-based COD Methods via Depth Anything V2' block of Table 1, VCP-DCN reports COD10K Fwβ = 0.930 (VMamba) and 0.932 (Swin). This is a ~0.09 gain over the same-backbone Samba row (0.842) and ~0.13 over DaCOD (0.795), while the accompanying improvements in Sm, Eϕ, and M are much smaller (e.g., Sm: 0.881 vs 0.870; M: 0.017 vs 0.019). The Fwβ value 0.930 also exceeds the reported CAMO/NC4K Fwβ (0.871) and is higher than the COD10K Sm (0.881), which is atypical for this benchmark. No ablation in Table 3 produces a gain larger than 0.03 Fwβ. These numbers are load-bearing for the 'consistently outperforms' claim, and the paper provides no error bars, multiple seeds, or runnable evaluation configuration to verify them. Please check for typographical errors or column misalignment and provide corrected or verified values.","section":"Table 1, Sec. 4.2"},{"comment":"The prototype contrastive losses L_MCL, L_MSL, L_RCL, and L_DCL compute logarithms of sums or differences of cosine similarities without clamping, a temperature, or a softmax formulation. Since cosine similarity lies in [-1,1], the argument of the logarithm (e.g., Cos(F_P_rc, F_P_dc) + Cos(B_P_r, B_P_d) in Eq. (3), or 1 - Cos(F_P_rs, F_P_ds) in Eq. (4)) can be zero or negative at initialization, making the loss undefined or infinite. The paper does not describe any numerical stabilization (e.g., epsilon or temperature). This is central to the SPE module, and the authors should clarify the exact implementation or provide a stable variant with a reference to the standard contrastive-learning form.","section":"Sec. 3.2, Eqs. (3)-(5)"},{"comment":"The code link is given in the abstract, but no commit hash, evaluation scripts, or training/validation configuration are provided. The numerical claims, especially the anomalous COD10K Fwβ values, cannot be independently reproduced without these artifacts. The authors should release the exact evaluation harness, the trained models or multiple random seeds, and the pseudo-depth generation pipeline. Without this, the central comparison table is not verifiable.","section":"Sec. 4.1 and Code Release"},{"comment":"All experiments rely on pseudo-depth from Depth Anything V2; the only depth-source comparison is DPT vs. Depth Anything V2 in Table 1. There is no sensitivity analysis to depth noise or to real sensor depth. Since the method's core contribution is a 'depth collaborative network,' it is important to show that the gains are not an artifact of the specific pseudo-depth estimator. Adding an experiment with noisy or alternative depth input (e.g., corrupting Depth Anything V2 outputs or using a lower-quality depth estimator) would directly address this robustness concern.","section":"Sec. 4.1 and Table 1 (depth-source comparison)"}],"minor_comments":[{"comment":"Several references are duplicated: [7] and [8] are the same paper; [10] and [11]; [23] and [24]; [50] and [51]. Please consolidate. This is particularly distracting for the authors' own prior work ([7]/[8]).","section":"References"},{"comment":"Ablation studies are reported only on CAMO. The title claims 'consistently outperforms' across three datasets; reporting ablations on at least COD10K or NC4K would strengthen that claim.","section":"Sec. 4.3"},{"comment":"The RGB-based group includes Samba, while the RGB-D-based DPT group also includes Samba. The grouping is confusing; please clarify whether Samba is used with RGB-only in one group and with depth in the other, or rename the rows to avoid ambiguity.","section":"Table 1"},{"comment":"Eq. (9) computes similarity scores only for the foreground modality-specific prototypes (F_P_rs, F_P_ds) against the consistency prototype. For symmetry with Eq. (10), where both foreground and background specific features are used in fusion, the paper should also define background-specific similarity scores or explain why only foreground prototypes are used in the decision-making mechanism.","section":"Sec. 3.4, Eq. (9)"},{"comment":"The caption says 'we visualize the semantic maps SM_r^3 and SM_r^3'; the second should presumably be SM_d^3 (depth branch).","section":"Fig. 7(a)"},{"comment":"Minor language issues: 'we presents VCP-DCN' in Sec. 5 and 'boots' in Sec. 4.3 ('F-MLA and B-MLA boots A1 by 9.8%') should be 'present' and 'boost', respectively.","section":"Sec. 5 and Sec 4.3"}],"recommendation":"major_revision","confidential_remarks":"The apparent COD10K Fwβ anomaly is the main barrier to acceptance. If the values are typographical, the paper's SOTA claim loses its strongest quantitative support on the largest benchmark; if they are genuine, the paper provides no explanation and no reproducible setup to verify them. The duplicated references, including the authors' own [7]/[8], should be cleaned up; this may be an oversight, but it also raises a citation-integrity concern that the editor may want to check. The architectural idea is interesting and the ablations are internally consistent, so I believe a major revision addressing the above points is the appropriate path."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First thing you should know: this is a solid incremental architecture paper, but its headline SOTA claim currently rests on a number that looks wrong. In Table 1, VCP-DCN reports COD10K Fwβ = 0.930 (VMamba) and 0.932 (Swin) with Depth Anything V2, versus 0.842 for Samba, the closest same-backbone comparison. No other metric on that dataset moves by more than a few hundredths: Sm goes 0.870 to 0.881, E 0.939 to 0.948, MAE 0.019 to 0.017. A +0.088 Fwβ jump with almost no change elsewhere is implausible. The stated gains on CAMO and NC4K are modest, around 0.008–0.01 Fwβ over Samba, which is believable and consistent with the ablations.\n\nWhat is genuinely new: the SPE/MDA/DAI pipeline with modality-consistency versus modality-specific prototype contrastive learning is a new assembly for COD, and the idea of preserving modality discriminability through the fusion stage is worth taking seriously. The ablations are internally consistent and the heatmap/t-SNE visualizations support the mechanism. The transfer to RGB-D SOD is a nice plus, and the paper is clearly written.\n\nSoft spots beyond the anomaly: duplicated references ([7]/[8], [10]/[11], [13]/[14], [23]/[24], [46]/[47], [50]/[51], [62]/[63]) suggest careless copy-paste. The percentage improvements cited in Sec 4.3 (e.g., 12.8% for DAI) do not match Table 4 (closer to 10%). No error bars or multiple seeds are reported. The RGB-D SOD split description is ambiguous. Reliance on Depth Anything V2 pseudo-depth is reasonable, but an experiment with real or corrupted depth would strengthen the claims. The code release lacks a commit hash and config files, which matters for verifying the 0.930 figure.\n\nIf that number is a typo, the \"consistently outperforms\" statement is weaker but still true on CAMO and NC4K. If it is not a typo, the paper needs to explain an outlier that no ablation accounts for. Either way, this is a serious, readable engineering contribution with a sound mechanism. I would send it to peer review rather than desk-reject it, with the clear expectation that the authors correct Table 1, supply runnable code, report variance, and clean up the references. I would not cite the SOTA claim until the anomaly is resolved, but the architecture itself is worth engaging with.","headline":"A coherent RGB-D COD architecture with a solid ablation story, but the headline SOTA claim rests on a COD10K Fwβ number that looks like a typo and needs verification before publication.","tokens_in":16559,"tokens_out":3468,"would_cite":true,"duration_ms":30022,"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":"The paper claims that a depth collaborative network, which explicitly learns modality-consistency and modality-specific prototypes via contrastive learning, achieves state-of-the-art camouflaged object detection by preventing RGB and depth","keywords":["camouflaged object detection","RGB-D fusion","multi-modality learning","prototype contrastive learning","depth estimation","visual state space model","image segmentation"],"falsifier":"Train the same pipeline on the same benchmarks with real depth maps (e.g., from a depth sensor) or with corrupted pseudo-depth and compare; if performance drops to RGB-only levels, the claimed depth benefit is an artifact of the specific estimator.","tokens_in":15556,"feed_emoji":"🎯","tokens_out":4465,"duration_ms":40363,"temperature":0.7,"pith_summary":"This paper addresses a core failure it identifies in existing RGB-D camouflaged object detection: when RGB and depth features are fused under a single mask loss, the modalities converge into homogeneous representations and lose their complementary strengths. The authors propose VCP-DCN, a depth collaborative network that progressively aligns, interacts, and fuses RGB and depth features while explicitly preserving what is shared and what is unique to each modality. They introduce prototype contrastive learning to the COD task, separating RGB and depth features into modality-consistency and modality-specific prototypes. If the claims hold, VCP-DCN becomes the new state of the art on three standard camouflaged-object benchmarks and also transfers to RGB-D salient object detection.","feed_headline":"Depth collaboration tops camouflaged-object benchmarks","feed_subtitle":"A prototype-contrastive network keeps RGB and depth cues separate, letting geometry reveal hidden objects.","key_machinery":"The central object is a set of learned prototype tokens—foreground and background, further separated into modality-consistency and modality-specific variants—derived from RGB and depth features via the SPE module. These prototypes are trained with four contrastive losses (modality-consistency, modality-specific, and two foreground-background separation losses). The MDA module uses the consistency prototypes to generate foreground and background masks, then applies masked linear attention to exchange cross-modal cues. The DAI module computes cosine similarities between specific and consistency prototypes to adaptively weight RGB and depth contributions during fusion. The entire mechanism is d","core_discovery":"VCP-DCN outperforms previous state-of-the-art COD methods on CAMO, COD10K, and NC4K, and it does so while using relatively few parameters and FLOPs. The central innovation is a progressive three-stage pipeline: the Separable Prototype Embedding (SPE) module uses prototype contrastive losses to separate modality-consistency and modality-specific prototypes for foreground and background; the Multi-modality Dual Attention (MDA) module exchanges cross-modal prototype-derived masks to refine features; and the Depth Adaptive Injection (DAI) module weighs RGB and depth contributions based on similarity between specific and consistency prototypes. The paper also shows the same network transfers to R","pith_inferences":["Editorial inference: the paper does not test robustness to depth errors; a natural extension would be to measure performance as pseudo-depth noise increases or with real depth sensors, to see if the depth benefit persists.","Editorial inference: the separable prototype space could serve as an interpretability tool, revealing what depth contributes beyond appearance for each camouflaged scene.","Editorial inference: the prototype contrastive mechanism may transfer to other modality pairs, such as RGB-thermal or RGB-multispectral, for concealed object detection in different domains."],"forward_implications":["RGB-D camouflaged object detection surpasses RGB-only methods on CAMO, COD10K, and NC4K, establishing depth as a reliable complementary cue for concealed objects.","Prototype contrastive learning prevents multimodal feature homogenization, directly addressing a bias the paper identifies in earlier RGB-D fusion approaches.","The same depth collaborative network transfers to RGB-D salient object detection, suggesting a general mechanism for RGB-D segmentation beyond camouflage.","With 60.2M parameters and 46.4G FLOPs, the architecture is efficient enough for real-time applications while maintaining high accuracy."],"fun_headline_variants":["Depth collaboration outperforms on three COD benchmarks","Prototype contrastive learning improves camouflaged object detection","Depth cues separate hidden objects from background","VCP-DCN: Depth collaboration reveals camouflaged objects","Depth-aware network beats SOTA in camouflaged object detection"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The results depend on the pseudo-depth maps generated by a monocular depth estimator being informative enough about geometry for camouflaged scenes; if those maps are noisy or uninformative, the depth branch contributes no signal.","fun_headline_variants_meta":{"raw":{"variants":["Depth collaboration outperforms on three COD benchmarks","Prototype contrastive learning improves camouflaged object detection","Depth cues separate hidden objects from background","VCP-DCN: Depth collaboration reveals camouflaged objects","Depth-aware network beats SOTA in camouflaged object detection"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000864,"raw_usage":{"total_tokens":3604,"prompt_tokens":783,"completion_tokens":2821,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":527,"completion_tokens_details":{"reasoning_tokens":2745}},"tokens_in":527,"tokens_out":2821,"duration_ms":18857,"temperature":1.0,"reasoning_tokens":2745,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T00:24:01.666249+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the same pipeline on the same benchmarks with real depth maps (e.g., from a depth sensor) or with corrupted pseudo-depth and compare; if performance drops to RGB-only levels, the claimed depth benefit is an artifact of the specific estimator.","supporting_citations":[],"review_version":1}