acceptodds
Withdrawn from ICLR 2027

Seed, Diffuse, and Refine: Reliability-Screened Prior Learning for Unsupervised Camouflaged Object Detection

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Unsupervised Camouflaged Object Detection (UCOD) aims to eliminate the dependence on costly pixel-level annotations, yet remains highly challenging due to the extremely low foreground-background contrast and the absence of reliable supervision. Recent UCOD approaches increasingly exploit priors from foundation models to alleviate annotation scarcity, but often directly propagate these priors without explicitly assessing their reliability, resulting in noisy pseudo supervision and limited robustness in complex camouflage scenarios. In this paper, we revisit UCOD from the perspective of prior reliability and propose a reliability-screened framework that selectively exploits foundation-model knowledge. Instead of directly inheriting model-generated priors, we leverage the rejection-aware capability of SAM3 to perform semantic probing and construct a SAM3-screened seed set for subsequent learning. Based on these screened but limited seeds, we introduce a diffusion-based segmentation head to capture the distributional characteristics of camouflage structures rather than directly fitting deterministic masks. Furthermore, we propose a SAM3-guided boundary refinement strategy, where diffusion predictions are converted into geometric prompts to adapt SAM3 for boundary-aware mask refinement. Finally, we develop a dual-sided distillation scheme that transfers complementary geometric and semantic priors to the final segmentation model. Extensive experiments on three COD benchmarks demonstrate that our method achieves state-of-the-art performance among UCOD methods.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Camouflaged Object Detection, Unsupervised Learning, SAM3

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_qrFqOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes a “Seed–Diffuse–Refine” unsupervised camouflaged object detection framework based on reliability screening: first, reliable seeds are screened using SAM3’s rejection mechanism; next, a diffusion segmentation head learns the camouflage structure; then, SAM3 is used for boundary refinement; and finally, the segmentation model is trained via geometric-semantic bilateral distillation. The framework achieved leading results on COD10K, CAMO, and NC4K.

Critical strengths

The author's line of reasoning is well-considered, but it overlooks an understanding of basic unsupervised tasks.

Critical weaknesses

Unsupervised tasks cannot rely on foundation models trained under fully supervised conditions, such as SAM3 and the PVT architecture. If that were possible, why not simply feed the data directly into SAM and have it generate the segmentation? The authors have completely overlooked the definition of “unsupervised”; processing can be based on foundation models like DINO, so I tend to reject this claim, and the rebuttal process cannot resolve this issue.

Questions for authors

Unsupervised tasks cannot rely on foundation models trained under fully supervised conditions, such as SAM3 and the PVT architecture. If that were possible, why not simply feed the data directly into SAM and have it generate the segmentation? The authors have completely overlooked the definition of “unsupervised”; processing can be based on foundation models like DINO, so I tend to reject this claim, and the rebuttal process cannot resolve this issue.

Reviewer_Vj1kOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper addresses unsupervised camouflaged object detection (UCOD) by explicitly screening and selectively exploiting priors from foundation models. The proposed Seed–Diffuse–Refine framework first exploits SAM3’s rejection-aware behavior and a semantic query pool to construct a compact but more reliable seed set. It then uses a conditional diffusion-based segmentation head with boundary perturbation to learn the distributional characteristics of camouflage structures. Subsequently, geometric prompts derived from diffusion predictions and SAM3 mask decoder fine-tuning enable boundary-aware mask refinement. Finally, dual-sided distillation transfers geometric priors across all samples while selectively retaining semantic supervision only for SAM3-screened samples.

Critical strengths
  • The motivation is reasonable. Pseudo-labels generated by foundation models may contain noise, and their quality can affect subsequent unsupervised learning. The paper focuses on screening and utilizing foundation-model priors, which is a meaningful research direction for reducing reliance on manual annotations in camouflaged object detection.
  • Technical novelty. Converting diffusion predictions into geometric prompts and fine-tuning only the SAM3 mask decoder for boundary refinement enables modular collaboration between generative structure learning and prompt-based refinement.
Critical weaknesses
  • Reliability screening lacks sufficient validation. SAM3 is trained on general images and may struggle to generalize to camouflaged scenes, so a non-empty mask alone does not establish that it has correctly identified the camouflaged target, particularly when the query pool includes the broad term “Animal.” Better performance against ground-truth annotations on accepted samples may reflect a preference for easier or more salient targets rather than an ability to assess pseudo-label reliability. This weakens the justification for the paper’s central mechanism.
  • Table 1 contains numerous errors and unfair comparisons. The values attributed to UCOS-DA actually correspond to the FOUND baseline reported in the UCOS-DA paper. The ProMaC values differ from those in the original paper, and SAM-DSA’s maximum E-measure is reported in the mean E-measure column. In addition, EASE, UCOD-DPL, and DualUCOD are represented only by their weaker DINOv1 configurations, while stronger published variants are omitted, limiting the fairness of the comparison. I remain skeptical about the rigor of the paper’s experimental evaluation.
  • The mechanism of semantic distillation remains unclear. The target is described as a “semantic embedding” on page 4 but as “semantic logits” on page 7, where MSE and L1 losses are applied directly between this target and the segmentation prediction. If the target is an embedding, its dimensionality and alignment with the prediction are unspecified. If it is a soft mask, it primarily provides soft supervision from the same SAM3 prediction, which is insufficient to establish that it supplies the claimed category-level semantic knowledge complementary to geometric supervision.
Questions for authors
  • The mechanism of semantic distillation remains unclear. The target is described as a "semantic embedding" on page 4 but as "semantic logits" on page 7, where MSE and L1 losses are applied directly between this target and the segmentation prediction. If the target is an embedding, its dimensionality and alignment with the prediction are unspecified. If it is a soft mask, it primarily provides soft supervision from the same SAM3 prediction, which is insufficient to establish that it supplies the claimed "category-level" semantic knowledge complementary to geometric supervision.
  • Please address the issues with Table 1 noted in Weakness 2 and explain the choice of baseline configurations. Given your use of SAM3 and PVTv2, please clarify why stronger published baselines, particularly EASE with SAM and DINOv2, are omitted. Please update the comparisons and reassess the conclusions accordingly.
  • Please clarify the form and dimensions of ZseedZ_{\text{seed}}, how it is aligned with the segmentation prediction, and why L1 and MSE losses are used. Please also explain how the existing ablations support the claim that semantic and geometric distillation provide complementary information.
Reviewer_tUHvOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper presents an unsupervised camouflaged object detection framework that screens SAM3 pseudo-labels, uses diffusion to learn object masks, refines them with SAM3, and distills the results into a final model. It is evaluated on CAMO, COD10K, and NC4K.

Critical strengths
  1. The paper addresses the practical problem of noisy pseudo-labels in unsupervised COD.
  2. The method is tested on three standard benchmarks and shows competitive results.
Critical weaknesses
  1. The comparisons are not comprehensive. Some recent relevant methods and stronger baseline settings are missing, which weakens the state-of-the-art claim.
  2. The novelty and contribution of each stage are not fully established. It is also unclear how diffusion inference works for samples without valid SAM3 seed masks.
Questions for authors
  1. Can the authors add more recent baselines and compare under more consistent settings?
  2. Can the authors clarify the inference procedure and better isolate the benefit of each component?

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

Related papers

Discussion (0)

Sign in to comment.