acceptodds
Withdrawn from ICLR 2027

BAPNet: Boundary and Ambiguity Priors for Interactive Camouflaged Object Segmentation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Camouflaged object segmentation involves strong foreground–background similarity and ambiguous boundaries, making pixel-level annotation costly and time-consuming. Interactive segmentation offers an efficient solution by using limited user interactions, such as clicks, to assist mask generation. However, existing methods are mainly designed for natural images and often struggle in camouflaged scenarios, where ambiguous boundaries and high prediction uncertainty can lead to inaccurate contours and foreground–background confusion. To address these challenges, we propose a Boundary and Ambiguity Priors for Interactive Camouflaged Object Segmentation (BAPNet) that jointly models boundary information and prediction uncertainty. Specifically, a Grassmann Boundary Guidance Module (GBGM) is introduced to explicitly predict object boundaries and guide the model to focus on foreground–background transition regions, while an Ambiguity-Prior Refinement Module (APRM) identifies potential areas with unreliable predictions and enhances local discrimination in ambiguous regions. To the best of our knowledge, BAPNet is the first systematic study of interactive camouflaged object segmentation. Extensive experiments on multiple camouflaged object datasets demonstrate that the proposed method achieves superior segmentation accuracy and interaction efficiency compared with existing approaches.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
interactive segmentation, camouflaged object detection, computer vision

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_B9dTOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes BAPNet, an interactive camouflaged object segmentation framework designed to improve segmentation accuracy and reduce annotation effort. To address ambiguous object boundaries and foreground–background confusion, BAPNet introduces two key components: a Grassmann Boundary Guidance Module (GBGM) that captures structural boundary information through local subspace comparisons, and an Ambiguity-Prior Refinement Module (APRM) that identifies potentially erroneous regions and refines features using ambiguity-guided attention. Experiments on four camouflaged object datasets (CAMO, CHAMELEON, COD10K, and NC4K) demonstrate that BAPNet consistently outperforms existing interactive segmentation methods in terms of the number of clicks required to achieve target IoU thresholds. Ablation studies further support the effectiveness of the proposed modules.

Critical strengths

1. The paper addresses a practical and challenging task of interactive camouflaged object segmentation with clear motivation.

2. Extensive experiments on four benchmarks demonstrate consistent improvements over existing methods, supported by ablation studies.

Critical weaknesses
  1. Limited Novelty: Boundary guidance and uncertainty-aware refinement have been widely explored in previous studies. The proposed modules appear to be incremental extensions of existing techniques.

  2. Insufficient Efficiency Analysis: The Grassmann-based eigendecomposition and multi-scale attention introduce additional computational overhead, but inference time, FLOPs, and memory consumption are not reported, despite the emphasis on interaction efficiency.

  3. Marginal Performance Improvements: The improvements over strong baselines such as MST are relatively small in several settings, and no statistical significance analysis is provided.

  4. Insufficient Validation of Key Modules: The paper lacks direct evaluations of boundary prediction quality and ambiguity estimation accuracy, making it difficult to verify the effectiveness of the proposed mechanisms.

  5. Limited Ablation Studies: The complementary benefits of GBGM and APRM are not always evident. Sensitivity analyses of key hyperparameters and comparisons with simpler feature refinement strategies are also missing.

  6. Incomplete Experimental Comparisons: Comparisons with representative prompt-based segmentation models such as SAM and SAM2 are missing. The evaluation also lacks real-world user studies to validate the claimed annotation time savings.

  7. Writing and Formatting Issues: The manuscript contains several typos and grammatical errors (e.g., “In Eq. equation 2”), inconsistent notation, and duplicate references (e.g., Fan et al., 2020a/2020b; Liu et al., 2022a/2022b).

Questions for authors
  1. Computational Efficiency: Could the authors provide inference time, FLOPs, and GPU memory comparisons with existing methods? How does the additional computational overhead of GBGM and APRM affect real-time interactive segmentation?

  2. Effectiveness of the Proposed Priors: Could the authors provide quantitative evaluations of boundary prediction accuracy and ambiguity estimation quality? How can we verify that the performance improvements actually result from the proposed boundary and ambiguity priors?

Reviewer_c3KZOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes BAPNet, an interactive segmentation framework specifically designed for camouflaged objects, where high foreground-background similarity and ambiguous boundaries make accurate segmentation challenging. The method introduces two main components: a Grassmann Boundary Guidance Module (GBGM), which exploits local feature subspace differences to enhance boundary estimation, and an Ambiguity-Prior Refinement Module (APRM), which identifies potentially erroneous predictions and refines ambiguous regions. The framework jointly incorporates boundary and ambiguity priors to improve segmentation accuracy and reduce corrective interactions. Experiments on four camouflaged object benchmarks (CAMO, CHAMELEON, COD10K, and NC4K) demonstrate improvements over existing interactive segmentation methods in terms of the number of clicks required to achieve target IoU thresholds. Ablation studies further evaluate the contributions of the two proposed modules and the Grassmann-based boundary estimation strategy.

Critical strengths
  1. Meaningful task motivation and targeted methodological design. The paper addresses the practical challenge of interactive segmentation in camouflaged scenes, where conventional interactive segmentation methods may struggle with indistinct object boundaries and foreground-background ambiguity. Explicitly incorporating boundary and ambiguity priors is well motivated by the characteristics of this task. The proposed framework provides a task-oriented approach to improving interactive annotation efficiency.

  2. Consistent experimental improvements and supporting ablation studies. The method is evaluated on four established camouflaged object datasets using multiple interaction-efficiency metrics, including NoC@85, NoC@90, and NoC@95. The reported results demonstrate improvements over the compared methods, while the ablation studies show that both GBGM and APRM contribute to performance. Additional comparisons with alternative boundary estimation strategies provide evidence supporting the effectiveness of the Grassmann-based design.

Critical weaknesses
  1. Limited methodological novelty. Although the proposed method is specifically designed for interactive camouflaged object segmentation, the overall framework largely follows the conventional interactive segmentation paradigm. The two proposed modules primarily enhance boundary representation and refine ambiguous regions, rather than introducing a fundamentally new interaction mechanism or segmentation formulation. While these designs are reasonable and task-oriented, the methodological advances appear relatively incremental, raising concerns about whether the technical novelty is sufficient for ICLR.

  2. Relatively modest performance improvements. Although BAPNet achieves improvements over existing methods, the performance gains are relatively limited, especially considering that the framework is specifically tailored to camouflaged object segmentation. For example, the ablation results show only marginal improvements at the stricter NoC@95 threshold on some datasets. Furthermore, the advantages of the Grassmann-based boundary estimation over simpler alternatives are not particularly substantial. These observations raise questions about whether the additional architectural complexity is justified by the achieved improvements.

Questions for authors
  1. Computational efficiency. The paper emphasizes interaction efficiency, but the computational overhead introduced by the proposed modules is not sufficiently analyzed. Could the authors provide comparisons of model parameters, FLOPs, and inference time per interaction against existing methods? This would help assess whether the reduction in corrective clicks justifies the additional computational complexity.

  2. Comparison with existing approaches for challenging visual conditions. Weak object boundaries, low foreground-background contrast, and ambiguous regions are common challenges in image segmentation, and various approaches have been proposed to address them. Could the authors further clarify how the proposed boundary and ambiguity modeling differs from these existing techniques, and provide additional comparisons or analysis to demonstrate its distinctive advantages? This would help establish the methodological novelty and general effectiveness of the proposed approach.

Reviewer_boxtOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

BAPNet addresses interactive camouflaged object segmentation by introducing two modules: GBGM for boundary guidance and APRM for ambiguity-aware refinement. Experiments on four camouflaged object datasets show improvements in click efficiency over existing methods, and ablation studies support the contributions of both modules.

Critical strengths

1.The paper addresses a meaningful problem and proposes two complementary modules to handle ambiguous boundaries and unreliable predictions in camouflaged scenes.

2.Experiments on four datasets show consistent improvements, while ablation studies provide evidence for the effectiveness of both modules.

Critical weaknesses
  1. The novelty of GBGM remains unclear, particularly whether its boundary prior provides supervision beyond the original segmentation masks.

  2. The annotation-time savings are not fully substantiated, and the lack of inference latency measurements makes the practical efficiency of BAPNet difficult to assess.

Questions for authors

1.Please report inference latency and computational overhead

2.How are the ground-truth boundary annotations obtained? Are they manually annotated or generated from the segmentation masks? Are the boundaries simply the edges of the target masks? If so, does the boundary prior provide additional supervision beyond the original mask supervision, or is it essentially another form of mask supervision?

3.How are the annotation-time reductions in Figure 1 calculated? Please provide the intermediate calculations and clarify whether these are estimated or measured annotation times.

4.Could the authors more clearly distinguish interactive camouflaged object segmentation from existing camouflaged object segmentation and annotation approaches, and further substantiate the novelty of the proposed task?

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

Related papers

Discussion (0)

Sign in to comment.