acceptodds
Under review as a conference paper at ICLR 2027

Learning When to Trust Geometry: Probe-Relative Source Preference for Camouflaged Object Detection

Abstract

Adaptive fusion assumes a gate can learn where one representation is more trustworthy than another. We test that assumption. We supervise a spatial gate with the detached pixelwise loss difference between two auxiliary segmentation probes—the exact minimizer of an entropy-regularized two-source risk—and instantiate it in \method, a camouflaged object detector fusing frozen SAM 2.1 semantics with frozen Depth Anything V2 geometry. is strong: after SHA-256 auditing removes seven contaminated training images, three seeds reach of 0.9402, 0.9174, 0.9264, and 0.9327 on CHAMELEON, CAMO, COD10K, and NC4K, and matched-protocol inference beats ZoomNeXt-B4 on all four datasets with non-overlapping paired bootstrap intervals in , , and . The gate nevertheless fails to identify trustworthy geometry: AUROC for the probe-preference event is 0.6057–0.6429, and as a selector it incurs 2.8–4.5 the regret of the trivial rule “always trust geometry.” We trace this to the target, not the gate. Because both probes are accurate, at 75.9–81.9% of pixels, which at confines the target to and leaves the gate a near-constant signal. Three instruments confirm the mechanism: the margin distribution; a five-condition corruption ladder in which mean gate response is perfectly rank-correlated with induced damage ( on every dataset) and therefore blind wherever damage is small; and oracles showing 25.4–33.7% of output-space complementarity unclaimed. Margin-based competence supervision degenerates precisely as its experts co-improve—a general obstacle for loss-difference routing, not an artifact of one architecture.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Object detection, segemtation, Gate interface, Camouflaged objection

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_hmycOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper studies whether a learned spatial gate can be taught where to trust one modality over another.

The question is instantiated in CLOAK, a camouflaged object detector that fuses two frozen modalities: SAM 2.1 semantics and Depth Anything V2 geometry. The gate is supervised by the detached pixelwise loss difference between two auxiliary segmentation probes. This signal is derived as the exact minimizer of an entropy-regularized two-source risk.

Experimental results show the detector itself is strong under a matched protocol, yet its gate fails to identify trustworthy geometry, performing even worse than the trivial constant rule “always trust geometry”. The authors attribute this to margin collapse: since both probes are accurate, ~80% of pixels have loss margins below 0.05 and receive near-constant targets at 0.5, leaving the gate little signal to learn from. They conclude that margin-based relative-competence supervision degenerates as its experts co-improve, a limitation they position as general to loss-difference routing.

Critical strengths

++ Using probes to diagnose when the gate works and when it fails is an interesting idea.

++ The paper provides extensive hyperparameter details, even listing the exact random seed numbers—something rarely seen in previous papers.

Critical weaknesses

-- The paper’s diagnosis is based on a single gating design, a single architecture, and a single task, yet it concludes that margin collapse is “a general obstacle for loss-difference routing, not an artifact of one architecture”. To support the generality claim, additional gating mechanisms and tasks should be analyzed.

-- Both branches come from the same RGB image—monocular depth is estimated from it too—so the two experts share the same information ceiling: as both probes improve, their errors shrink together and the margin collapses almost by construction. The observed failure may thus be a property of same-source fusion, not loss-difference routing in general. With genuinely different inputs (e.g., RGB plus sensor depth or thermal), one expert can be right where the other is wrong, so margins can stay wide even when both are accurate. The authors should test such a setting to see whether collapse still occurs.

-- The gate is supervised by targets derived from probe loss differences, not by the final task loss alone. A gate trained end-to-end from the task loss could learn a different routing mechanism from one shaped by this proxy supervision, so conclusions drawn under probe supervision may not transfer to the setting the paper claims to generalize to. A simple ablation—removing the probe supervision and letting the gate learn from the task loss—would settle this, but it is absent.

-- The paper suggests fixes for gate unreliability, but none is ever tested. Validation stops at discussion, with no experiment showing that any remedy actually prevents margin collapse or restores routing.

-- The paper frames its contribution as learning when to trust geometry, yet the proposed gate never actually learns a reliable “when”: it collapses to over- or under-trusting, its confidence tracks neither depth correctness nor causal benefit, and no working mechanism is ultimately delivered. What remains is a diagnosis of failure, not a solution to the stated problem—so the central claim of the title overstates the contribution.

-- The method is illustrated only by a simple text-block diagram (Fig. 1). The main text contains no visualizations of gate maps, margin distribution, or corruption responses. In addition, a stray vertical black edge remains along the left side of Fig. 1, suggesting a screenshot was pasted without cropping.

-- Every appendix reference renders as “??” (L067, L143, L198, L209, L358, L417, L450, L453), and in-text table references are similarly broken (e.g., L267–269, L415).

Questions for authors

N/A

Reviewer_V7UXOpenReview ↗
Rating
2
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

The paper poses a question: can a gate actually be taught where one source is more trustworthy than another? As a testbed, the authors build a camouflaged object detector fusing two types of features, supervising the gate with the loss difference between two auxiliary probes. Its contribution is diagnostic: inside a demonstrably strong detector, the authors show that the gate cannot select which source is trustworthy, and they explain why, validated through a coherent set of instruments. This constitutes the paper's main methodological value. The authors further claim that this failure is generic to loss-difference routing (e.g., MoE, multimodal fusion), but this generalization is asserted rather than validated.

Critical strengths

The paper's core strength is that it goes beyond observing a failure and delivers a mechanism-level diagnosis with strong experimental discipline. It also provides a logically coherent validation methodology.

Critical weaknesses
  1. The paper presents margin collapse as a general obstacle for loss-difference routing in MoE and multimodal fusion, but validates it only within a single architecture and a single modality pair, so the claim currently exceeds the evidence.
  2. The paper localizes the cause of the failure but tests none of the remedies it proposes. If any of these remedies are easy to implement, verifying them experimentally would substantially strengthen the paper.
  3. The cross-attention baseline used for comparison has no architectural description anywhere in the paper, so the fairness of this comparison cannot be assessed.
Questions for authors
  1. Can you provide at least one experiment that closes the causal loop of the margin-collapse diagnosis? (See Weakness 2.)
  2. Can you substantiate the generalization claim, or appropriately scope it down?
  3. Please describe the CA (cross-attention) baseline.
  4. How was τ=0.25 chosen? How sensitive are the margin-collapse statistics to τ ∈ {0.1, 0.5, 1.0}?
  5. Will you release the code, evaluation scripts, and audit artifacts?
Reviewer_pzkhOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper investigates the reliability of adaptive fusion mechanisms for camouflaged object detection and proposes CLOAK, a framework combining semantic features from SAM 2.1 and geometric features from Depth Anything V2. The authors introduce a probe-relative source preference strategy, where the pixel-wise loss difference between two auxiliary segmentation probes is used to supervise a spatial gating mechanism. Interestingly, although the proposed framework achieves promising segmentation performance, the authors find that the learned gate struggles to identify when geometric information should be trusted. They further analyse this limitation and attribute it to the diminishing loss differences between increasingly accurate experts, which results in less informative supervision signals. Extensive experiments are conducted on four benchmark datasets. The results demonstrate the competitive segmentation performance of CLOAK while revealing the limitations of loss-difference-based routing strategies.

Critical strengths
  1. The paper investigates an important but relatively underexplored question in adaptive multimodal fusion: whether a learned gating mechanism can effectively distinguish reliable information from different modalities. This provides an interesting perspective beyond conventional segmentation performance improvements.

  2. The authors provide a mathematical formulation of probe-relative source preference and analyse the limitations of loss-difference-based supervision. The discussion of margin collapse offers useful insights into why adaptive gating mechanisms may become less effective as individual experts improve.

  3. The authors conduct experiments on four benchmark datasets and provide extensive ablation studies, corruption experiments, and diagnostic evaluations. These results support the main observations and provide useful empirical evidence for the limitations discussed in the paper.

Critical weaknesses
  1. Although the motivation is interesting, the proposed CLOAK framework mainly integrates existing pretrained models, including SAM 2.1 and Depth Anything V2, with a relatively conventional gating-based fusion mechanism. The main contribution lies in the analysis of the limitations of probe-relative supervision rather than the development of a fundamentally new fusion architecture. The authors should clarify the methodological novelty.
  2. According to Table 4, the proposed probe-relative supervision does not consistently outperform simpler alternatives, such as mask-based supervision or ungated fusion. In addition, the learned gate exhibits relatively weak discrimination ability, with AUROC values ranging from 0.6057 to 0.6429. It raises questions about the practical value of the proposed supervision strategy.
  3. Although the authors compare CLOAK with ZoomNeXt-B4 under a matched inference protocol, the quantitative comparisons with other recent geometry-assisted methods remain relatively limited.
  4. Although Figure 1 illustrates the overall framework, it is relatively abstract and does not clearly highlight the key motivation and contributions of this work.
Questions for authors

Please refer to the Critical Weaknesses section.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.