acceptodds
Under review as a conference paper at ICLR 2027

Discriminative Evidence Mining and Reconstruction for Efficient Camouflaged Object Detection

Abstract

Camouflaged Object Detection (COD) aims to detect objects that closely resemble their background in complex scenes. Existing lightweight COD methods typically reduce computation by compressing feature dimensions and simplifying feature interactions. However, such approaches tend to weaken the subtle discriminative evidence required to distinguish camouflaged objects from their backgrounds, including semantic cues and fine-grained structural details. Therefore, we propose an efficient Discriminative Evidence Mining and Reconstruction Network (DMRNet), which incorporates the Discriminative Evidence Mining (DEM) strategy and the Discriminative Evidence Reconstruction (DER) strategy. Through the DEM, DMRNet captures cross-layer feature discrepancies, with foreground-background competition guiding evidence selection and background interference suppression reducing misleading responses, thereby obtaining more reliable discriminative evidence. Through the DER, DMRNet uses the obtained evidence to guide hierarchical feature reconstruction, thereby enhancing semantic representations at higher levels and fine-grained structural details at the shallow level. DMRNet achieves competitive detection performance on four benchmark datasets while maintaining a low parameter count and computational cost, thereby achieving an effective balance between accuracy and efficiency. The source code will be released upon publication of the paper.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Camouflaged Object Detection, Image Segmentation, Lightweight Network

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_YL7COpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The authors propose an efficient Discriminative Evidence Mining and Reconstruction Network (DMRNet) that integrates a Discriminative Evidence Mining (DEM) strategy and a Discriminative Evidence Reconstruction (DER) strategy. Specifically, DEM captures cross-layer feature discrepancies and exploits foreground–background competition to guide evidence selection while suppressing background interference, thereby attenuating misleading responses and yielding more reliable discriminative evidence.

Critical strengths
  1. The paper addresses the degradation of discriminative evidence caused by feature compression in lightweight camouflaged object detection (COD). The research problem is clearly defined. To address this issue, the proposed DMRNet introduces effective and innovative modules, achieving the best performance among the compared methods while maintaining computational efficiency. Therefore, the work has practical significance.

  2. The paper conducts extensive comparisons with a large number of competing methods on four benchmark datasets and achieves state of the art performance. In addition, the hierarchical feature visualizations and component-wise efficiency analyses provide more comprehensive evidence than numerical metrics alone, making the final results more convincing.

Critical weaknesses
  1. In the ablation study, introducing CDEM leads to a substantial performance improvement. However, it remains unclear whether this gain is mainly attributable to the decoder structure itself or to the proposed foreground-background competition mechanism. Moreover, removing MFC results in only a marginal performance drop. More detailed and carefully controlled ablation studies are therefore required to verify these points.

  2. The experimental setup should provide additional implementation details, including the validation set partition, to ensure the reproducibility of the reported results.

  3. Figure 2 does not clearly demonstrate the advantages of the proposed method; the authors should revise it to better highlight these advantages.

Questions for authors

The authors are asked to address the weaknesses I have raised and provide as many reasonable explanations as possible.

3 comments
AuthorsOpenReview ↗

Response to Weakness (1) from Reviewer YL7C

Response to the Only Backbone Setting and the Ablation Results of MFC

Thank you for pointing out the issues regarding the performance improvement and the MFC ablation study. We provide our explanations below.

First, the substantial performance improvement after introducing CDEM is attributable to the differences in feature processing between the two settings. Only Backbone directly feeds the four hierarchical feature maps output by SMT-T into 1×11\times1 convolutions for prediction, without performing any COD-specific feature learning. Moreover, SMT-T itself is not specifically designed for COD, resulting in limited detection performance. This setting is designed to demonstrate that the performance of our model stems from the proposed architectural design rather than relying solely on the backbone.

Second, regarding the relatively limited performance degradation after removing MFC, we believe this should be understood in the context of MFC's role in the overall architecture. MFC is not designed to independently undertake the COD task. Instead, it further enhances the discriminative capability of features based on existing hierarchical feature representations through prediction-guided foreground-background prototype competition and uncertainty modeling. Therefore, removing MFC does not deprive the model of its basic feature representation and detection capabilities, but weakens its explicit modeling of foreground-background differences and adaptive feature modulation, consequently affecting the final detection performance. Furthermore, DMRNet targets lightweight COD, where the computational budget is strictly limited. Under such limited computational resources, every performance improvement is challenging to achieve.

To further verify the individual contributions of the mechanisms within MFC, we conducted three additional ablation experiments by removing foreground-background prototype competition (w/o F/B C), confidence weighting (w/o CW), and uncertainty-guided weighting (w/o UG), respectively, while keeping the other network structures and experimental settings unchanged. The experimental results are presented in the following table.

SettingsCAMOCHAMELEONCOD10KNC4K
Sm↑S_m\uparrowFβw↑F_\beta^w\uparrowM↓M\downarrowEm↑E_m\uparrowSm↑S_m\uparrowFβw↑F_\beta^w\uparrowM↓M\downarrowEm↑E_m\uparrowSm↑S_m\uparrowFβw↑F_\beta^w\uparrowM↓M\downarrowEm↑E_m\uparrowSm↑S_m\uparrowFβw↑F_\beta^w\uparrowM↓M\downarrowEm↑E_m\uparrow
w/o F/B C0.8790.8380.0460.9310.9160.8700.0210.9610.8730.7940.0220.9350.8930.8480.0310.938
w/o CW0.8780.8380.0460.9290.9130.8690.0210.9600.8720.7950.0220.9350.8920.8470.0310.937
w/o UG0.8790.8410.0460.9310.9150.8710.0210.9600.8710.7930.0230.9330.8920.8490.0320.936
DMRNet0.8840.8460.0440.9330.9180.8750.0190.9620.8730.7960.0220.9360.8930.8490.0310.938

As shown in the table, removing any of the three mechanisms degrades performance, confirming their respective contributions. Among them, CW contributes most by improving prototype reliability, which supports subsequent foreground-background competition and discriminative evidence mining. F/B C provides discriminative cues through competition between foreground and background prototypes, while UG adjusts spatial routing weights according to prediction uncertainty. These mechanisms therefore serve complementary roles in prototype construction, foreground-background discrimination, and adaptive feature routing. Their individual contributions are further validated by the ablation results.

Following your suggestion, we have added the above three ablation experiments, complete experimental results, and corresponding analyses to the Appendix 10 to further validate the roles of the individual mechanisms within MFC.

AuthorsOpenReview ↗

Response to Weakness (2) and Weakness (3) from Reviewer YL7C

Response to the Experimental Settings and the Method Architecture Figure

Thank you for pointing out the insufficient experimental details and the lack of clarity in the method architecture figure. We have made corresponding revisions based on your suggestions.

(1) Regarding the experimental details. We clarify that no separate validation set is used in our experiments. Following the checkpoint selection setting adopted by most methods in Table 1 (e.g., PRNet, KCNet, UTNet, and CHNet), we evaluate the checkpoint from each epoch on the CAMO test set and select the one with the lowest MAE. We have added these details to Section 4.1 (Implementation Details) of the revised manuscript to clarify the model selection procedure and improve reproducibility.

(2) Regarding the presentation of Fig. 2. Following your suggestion, we have redrawn Fig. 2 to provide a clearer illustration of the overall architecture of DMRNet. We have also further elaborated on the internal operations and feature interactions of the three components in CDEM, i.e., FP, MFC, and CEF, making the design rationale and overall workflow of the proposed method more intuitive and easier to understand.

AuthorsOpenReview ↗

Request Feedback from Reviewer YL7C

Dear Reviewer YL7C, 

Thank you again for your thoughtful comments on our submission. As the discussion period is coming to a close, we would like to ask whether our responses have addressed your concerns. If any points remain unclear or if you have additional questions, we would greatly appreciate your guidance and will be happy to clarify further. If you feel that our revisions and explanations have resolved the issues you raised, we would be grateful if you could consider updating your evaluation accordingly. 

We sincerely look forward to your feedback, and thank you for your time and effort throughout this process. 

Best regards, 

Authors of submission36108

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

This paper proposes the Discriminative Evidence Mining and Reconstruction Network (DMRNet) to address the limited feature representation capacity and insufficient foreground-background discrimination in lightweight Camouflaged Object Detection (COD). The proposed method consists of two main components: (1) Discriminative Evidence Mining (DEM), which enhances object-related features through cross-layer feature interactions, foreground-background prototype competition, and background interference suppression; and (2) Discriminative Evidence Reconstruction (DER), which employs lightweight and grouped convolutions to enhance high-level semantic representations and low-level fine-grained structural details. Experiments on four public COD benchmarks demonstrate that DMRNet achieves competitive detection performance while maintaining relatively low parameter counts and computational costs.

Critical strengths
  1. The research problem is practically meaningful. The paper addresses the degradation of discriminative information caused by computational simplification in lightweight COD models. Balancing detection accuracy and inference efficiency is an important problem with practical relevance.

  2. The overall network architecture is reasonably designed. The proposed framework integrates cross-layer feature fusion, foreground-background competition, background interference suppression, and hierarchical feature reconstruction. The roles of the individual components are clearly defined, resulting in a relatively coherent architecture.

  3. The experimental results are competitive. The authors conduct evaluations on four public benchmarks, including CAMO, CHAMELEON, COD10K, and NC4K, and compare DMRNet with various existing methods. With only 12.46M parameters and 9.15G FLOPs, DMRNet achieves 53.74 FPS and competitive performance across multiple evaluation metrics. The paper also provides ablation studies on the main components and their submodules.

Critical weaknesses
  1. The novelty of the core method is limited, and direct comparisons with classical attention mechanisms are missing. The proposed DEM, particularly CDEM, can essentially be interpreted as an attention-based feature modulation mechanism guided by spatial localization information from high-level predictions. Foreground-background prototype construction, cosine similarity computation, uncertainty-based weighting, and cross-layer feature fusion are all established techniques. Although the authors integrate these operations into a unified framework, the unique advantages of this design over existing attention mechanisms remain unclear. In particular, the paper does not provide direct comparisons between CDEM and classical attention modules such as CBAM and SENet. Therefore, the current experimental evidence is insufficient to demonstrate that the proposed complex design is more effective than simpler, established attention mechanisms.

  2. The key formulations are largely heuristic and lack sufficient theoretical motivation and experimental justification. Equations (7)–(9) construct feature routing weights by combining prototype similarity differences, a hyperbolic tangent function, prediction uncertainty, linear weighting, and clipping operations. However, the authors do not adequately explain why this particular formulation is adopted. Specifically, Equation (8) uses a quadratic function to estimate prediction uncertainty, but its advantages over alternative uncertainty measures, such as predictive entropy, remain unclear. Furthermore, the linear combination and clipping operation in Equation (9) are not compared with alternative fusion strategies. Additional theoretical justification and ablation studies are needed to establish the necessity and effectiveness of these design choices.

Questions for authors
  1. Novelty and comparison with classical attention mechanisms. The proposed CDEM resembles an attention-based feature modulation mechanism guided by high-level predictions. Could the authors clarify the key methodological differences between CDEM and existing attention mechanisms such as CBAM and SENet? Are there any controlled comparisons demonstrating that the proposed foreground-background competition mechanism provides advantages over these simpler alternatives under comparable computational budgets?

  2. Motivation for the heuristic formulations. Could the authors further justify the design choices in Equations (7)–(9), particularly the use of the hyperbolic tangent function, the quadratic uncertainty measure, and the linear combination with clipping? Is there any theoretical or empirical evidence supporting these choices over alternative formulations?

  3. Parameter sensitivity and robustness. How sensitive is the proposed method to the temperature coefficient, clipping bounds, and learnable scaling factors in Equations (7), (9), and (10)? Have the authors observed significant performance variations under different parameter settings or across different datasets? Clarifying the selection procedure and stability of these parameters would help assess the robustness of the proposed method.

  4. Fairness of the baseline and effectiveness of the core mechanisms. Table 2 reports a substantial improvement in CAMO S-measure from 0.551 to 0.874 after introducing CDEM. Could the authors clarify whether the backbone-only baseline includes an equivalent multi-scale decoding and feature fusion structure? If not, how can the improvement be attributed specifically to the proposed discriminative evidence mining mechanism rather than the additional decoding capacity? Furthermore, can the authors provide existing ablation results or analyses that separately quantify the contributions of foreground-background prototype competition and uncertainty guidance?

7 comments
AuthorsOpenReview ↗

Response to Weakness (1) and Question (1) from Reviewer vVfc

(1) Response to the Novelty of DMRNet

Thank you for your comments on the novelty of our work. However, the specific design, functionality, and collaboration mechanisms of some operations in our model differ from those of existing methods. We offer the following explanation.

  • First, using high-level predictions to provide spatial localization cues is common in COD. However, unlike existing methods that convert them into spatial attention maps to enhance corresponding features, we use high-level predictions and their confidence maps to construct foreground and background prototypes for discriminative feature mining.

  • Furthermore, CDEM is not a conventional attention mechanism. Although CDEM employs element-wise modulation, its modulation weights are not learned through channel or spatial attention, but are jointly determined by foreground-background prototype competition responses and prediction uncertainty. Therefore, CDEM differs fundamentally from classical attention mechanisms in both weight generation and methodological mechanisms. Specifically, prototype competition captures the foreground-background tendency of features, while uncertainty-guided weighting adaptively adjusts spatial routing weights according to prediction uncertainty. These two mechanisms jointly regulate cross-layer information propagation. We do not claim uncertainty measurement itself as an independent contribution, nor do we consider cosine similarity an independent contribution, as it merely serves as a computational tool for measuring feature-prototype relationships.

  • Moreover, although cross-level feature fusion is common in COD, it is not the innovation we aim to emphasize. Our contribution lies in designing a distinct and lightweight cross-level interaction scheme that balances model complexity and detection performance, achieving favorable computational efficiency and inference speed.

Therefore, the novelty of our work does not lie in introducing entirely new basic operations, but in leveraging confidence-guided dual-prototype competition and uncertainty-aware routing to achieve more targeted discriminative evidence mining and cross-layer information propagation.

(2) A Response to the Comparison Between CDEM and Classical Attention Mechanisms

Thank you for suggesting a comparison between CDEM and classical attention mechanisms. To further verify the effectiveness of CDEM, we replaced it with CBAM and SENet, respectively, while keeping the settings of all other modules unchanged. Specifically, we concatenated the projected low-level and high-level features, fed them into CBAM or SENet, respectively, and adopted residual connections for the outputs. The experimental results are presented in the following table.

SettingsFLOPs (G)Params (M)CAMOCHAMELEONCOD10KNC4K
Sm↑S_m\uparrowFβw↑F_\beta^w\uparrowM↓M\downarrowEm↑E_m\uparrowSm↑S_m\uparrowFβw↑F_\beta^w\uparrowM↓M\downarrowEm↑E_m\uparrowSm↑S_m\uparrowFβw↑F_\beta^w\uparrowM↓M\downarrowEm↑E_m\uparrowSm↑S_m\uparrowFβw↑F_\beta^w\uparrowM↓M\downarrowEm↑E_m\uparrow
SENet-based9.4613.000.8760.8310.0470.9270.9120.8700.0210.9600.8710.7940.0220.9330.8900.8470.0310.936
CBAM-based9.4813.010.8740.8300.0480.9250.9140.8700.0200.9620.8720.7940.0220.9350.8910.8470.0310.936
CDEM-based9.1512.460.8840.8460.0440.9330.9180.8750.0190.9620.8730.7960.0220.9360.8930.8490.0310.938

As shown in the table, CDEM achieves the best overall detection performance, demonstrating more effective discriminative evidence mining than classical channel and spatial attention mechanisms. Moreover, with only 12.46M parameters and 9.15G FLOPs, CDEM achieves better detection performance with lower computational cost and fewer parameters than the CBAM- and SENet-based replacements.

Following your suggestion, we have added the implementation details, experimental results, and corresponding analyses of these replacement experiments to Appendix 9 of the revised manuscript.

AuthorsOpenReview ↗

Response to Weakness (2) and Question (2) from Reviewer vVfc

Response to the Design Choices in Eqs. (7)–(9)

Thank you for pointing out the need for further clarification of the design choices in Eqs. (7)–(9). Therefore, we have provided additional explanations from the perspectives of function properties and experimental validation.

(1) Motivation behind the design choices in Eqs. (7)–(9).

  • Hyperbolic tangent function. Eq. (7) maps the similarity difference between foreground and background prototypes into a bounded response in [-1,1]. Its monotonicity preserves the direction and relative magnitude of the similarity difference, while the temperature coefficient τ\tau controls mapping sensitivity.

  • Quadratic uncertainty measure. Eq. (8) uses Ui=4Mi(1−Mi)U_i=4M_i(1-M_i) to measure prediction uncertainty, which reaches 0 at probabilities of 0 or 1 and its maximum of 1 at 0.5, providing a continuous and symmetric measure of prediction ambiguity.

  • Linear combination and clipping operation. Eq. (9) combines prototype competition and prediction uncertainty through learnable coefficients, with a constant term of 1 providing a neutral baseline when both contributions are zero. Clipping constrains routing weights to prevent excessive feature enhancement or suppression.

(2) Ablation experiments with alternative formulations.

To verify these design choices, we conduct four ablation experiments, i.e., w/o Tanh removes the hyperbolic tangent in Eq. (7), w/o Clip removes clipping in Eq. (9), Entropy replaces the quadratic uncertainty measure in Eq. (8) with entropy, and Multiplicative replaces the additive combination in Eq. (9) with a multiplicative one. Each experiment modifies only the corresponding operation, keeping all other settings unchanged.

SettingsCAMOCHAMELEONCOD10KNC4K
Sm↑S_m\uparrowFβw↑F_\beta^w\uparrowM↓M\downarrowEm↑E_m\uparrowSm↑S_m\uparrowFβw↑F_\beta^w\uparrowM↓M\downarrowEm↑E_m\uparrowSm↑S_m\uparrowFβw↑F_\beta^w\uparrowM↓M\downarrowEm↑E_m\uparrowSm↑S_m\uparrowFβw↑F_\beta^w\uparrowM↓M\downarrowEm↑E_m\uparrow
w/o Tanh0.8720.8300.0470.9250.9120.8700.0210.9550.8710.7920.0220.9320.8910.8460.0310.935
w/o Clip0.8760.8300.0460.9290.9140.8680.0220.9590.8700.7890.0230.9320.8920.8480.0310.936
Entropy0.8800.8380.0450.9310.9140.8720.0200.9610.8710.7930.0220.9340.8920.8480.0310.938
Multiplicative0.8760.8330.0470.9280.9160.8740.0190.9600.8720.7940.0220.9330.8910.8470.0310.936
DMRNet0.8840.8460.0440.9330.9180.8750.0190.9620.8730.7960.0220.9360.8930.8490.0310.938

As shown in the table, the complete model achieves the best or tied-best results across all 16 metrics. Removing Tanh or Clip degrades performance, confirming the benefits of bounded competitive responses and constrained routing weights. Replacing the quadratic uncertainty measure with entropy or the additive combination with a multiplicative one also reduces performance, validating the effectiveness of the adopted formulations.

Following your suggestion, we have added the motivations behind the design choices in Eqs. (7)–(9), together with the experimental settings, results, and corresponding analyses of these alternative formulations, to Appendix 11 of the revised manuscript.

AuthorsOpenReview ↗

Response to Question (3) from Reviewer vVfc

Response to the Parameter Sensitivity Analysis

Thank you for your comments. Following your suggestions, we conducted further experiments and analyses on the relevant parameters.

(1) Temperature coefficient τ\tau in Eq. (7). Appendix 5 originally reported average performance across the four datasets. We have added dataset-wise results under different temperature settings and corresponding analyses.

Table 1. Sensitivity analysis of the temperature coefficient τ\tau on four benchmark datasets.

SettingsCAMOCHAMELEONCOD10KNC4K
Sm↑S_m\uparrowFβw↑F_\beta^w\uparrowM↓M\downarrowEm↑E_m\uparrowSm↑S_m\uparrowFβw↑F_\beta^w\uparrowM↓M\downarrowEm↑E_m\uparrowSm↑S_m\uparrowFβw↑F_\beta^w\uparrowM↓M\downarrowEm↑E_m\uparrowSm↑S_m\uparrowFβw↑F_\beta^w\uparrowM↓M\downarrowEm↑E_m\uparrow
τ=1\tau=10.8720.8270.0500.9210.9130.8670.0220.9540.8710.7920.0230.9320.8910.8460.0320.935
τ=2\tau=20.8800.8360.0460.9290.9110.8710.0210.9590.8710.7900.0230.9320.8920.8480.0300.936
τ=3\tau=30.8770.8400.0470.9290.9160.8700.0200.9590.8700.7930.0230.9310.8920.8490.0310.936
τ=4\tau=40.8840.8460.0440.9330.9180.8750.0190.9620.8730.7960.0220.9360.8930.8490.0310.938
τ=5\tau=50.8760.8310.0470.9250.9150.8750.0210.9640.8690.7870.0230.9300.8900.8440.0320.934
τ=6\tau=60.8740.8290.0480.9250.9130.8670.0210.9620.8660.7820.0230.9280.8900.8430.0320.934

As shown in Table 1, CAMO is sensitive to τ\tau. Performance improves up to τ=4\tau=4 and declines beyond it. CHAMELEON and NC4K remain relatively stable. On COD10K, results change little for τ=1\tau=1–3, but FβwF_\beta^w and EmE_m decline when τ>4\tau>4. Overall, τ=4\tau=4 provides good performance across the four datasets and is retained as the default setting.

(2) Learnable parameters αi\alpha_i, βi\beta_i, and γi\gamma_i in Eqs. (9)–(10). These are learnable rather than manually fixed hyperparameters. αi\alpha_i and βi\beta_i control the contributions of foreground-background competition and prediction uncertainty to routing weights, respectively, while γi\gamma_i controls feature modulation strength. They adapt during training, so manual sensitivity analysis is unnecessary.

(3) Clipping bounds rmin⁡r_{\min} and rmax⁡r_{\max} in Eq. (9). We added comparative experiments and retrained and evaluated the model under each setting.

Table 2. Sensitivity analysis of the routing weight clipping bounds on four benchmark datasets.

SettingsCAMOCHAMELEONCOD10KNC4K
Sm↑S_m\uparrowFβw↑F_\beta^w\uparrowM↓M\downarrowEm↑E_m\uparrowSm↑S_m\uparrowFβw↑F_\beta^w\uparrowM↓M\downarrowEm↑E_m\uparrowSm↑S_m\uparrowFβw↑F_\beta^w\uparrowM↓M\downarrowEm↑E_m\uparrowSm↑S_m\uparrowFβw↑F_\beta^w\uparrowM↓M\downarrowEm↑E_m\uparrow
(0.5, 1.5)0.8780.8350.0450.9320.9060.8580.0230.9520.8650.7810.0240.9270.8880.8380.0330.932
(0.25, 2.0)0.8840.8460.0440.9330.9180.8750.0190.9620.8730.7960.0220.9360.8930.8490.0310.938
(0.1, 3.0)0.8790.8370.0470.9290.9130.8680.0220.9570.8730.7940.0230.9340.8930.8490.0310.938

As shown in Table 2, the default bounds (0.25, 2.0) achieve the best or tied-best results across all 16 metrics. Narrower bounds restrict adaptive feature modulation, while wider bounds may cause imbalanced feature responses and weaken the balance between preserving useful information and suppressing background interference. Thus, the default bounds provide a better balance.

Following your suggestion, we have revised the manuscript accordingly. Appendix 5 includes the temperature sensitivity results and analyses for each dataset, while Appendix 6 presents the comparisons and analyses of different clipping bounds.

Reviewer_v72EOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes DMRNet, a lightweight network for camouflaged object detection. It combines discriminative evidence mining, using prediction-guided foreground-background prototypes and feature modulation, with hierarchical feature refinement to enhance semantic and structural information. Experiments on four benchmarks show competitive detection performance with 12.46M parameters and 9.15G FLOPs.

Critical strengths

(1) The paper is clearly written and easy to follow and the experiments include comparisons with recent COD models.

(2) The qualitative results clearly illustrate improvements in object completeness and background suppression.

Critical weaknesses

(1) The proposed modules largely build on established foreground/background modeling, feature reweighting, and progressive refinement.

  • For CDEM component, foreground/background prototypes and similarity matching have been explored in PANet [1], while prediction-guided foreground/background modeling and refinement are established in PFNet [2] and CamoFormer [3]. The additional contribution of confidence-weighted prototype competition is not sufficiently demonstrated.
  • For BIS component, the equation (13) is a learned feature gate, while distraction suppression is already central to PFNet [2]. The paper does not clearly distinguish BIS from conventional reweighting or demonstrate that its gate specifically captures background interference. Thus, the proposed combination is effective, but its methodological contribution beyond established foreground/background modeling and progressive decoding remains insufficiently demonstrated.

(2) The ablation removes MFC as a whole in Table 3 but does not isolate its key design choices: confidence-weighted prototype extraction and foreground-background prototype competition. It therefore remains unclear which of these choices drives the improvement and whether the full formulation is necessary.

[1] Panet: Few-shot image semantic segmentation with prototype alignment[C]//2019 IEEE/CVF International Conference on Computer Vision (ICCV). IEEE, 2019: 9196-9205.

[2] Camouflaged object segmentation with distraction mining[J]. arXiv preprint arXiv:2104.10475, 2021.

[3] Camoformer: Masked separable attention for camouflaged object detection[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024, 46(12): 10362-10374.

Questions for authors

Please see weakness

8 comments
AuthorsOpenReview ↗

Response to Weakness (1) from Reviewer v72E

(1) Response to the Differences Between DMRNet and Existing Methods

Thank you for your comments on the novelty of our work. However, there are some differences between our model and the models you mentioned. We provide the following clarifications.

  • First, regarding the difference from PANet. PANet adopts a Support-Query paradigm, constructing prototypes from annotated masks and performing pixel-level classification of query features through prototype matching. In contrast, we construct foreground and background prototypes using high-level predictions from the same image and employ them for discriminative evidence mining at the current level and cross-layer information transfer.

  • Second, regarding foreground-background modeling. PFNet explores foreground and background information through separate context branches, while CamoFormer employs prediction-mask-guided foreground and background attention branches. In contrast, we use prediction-guided prototype competition to guide cross-layer discriminative evidence mining and transfer. Therefore, our method differs from these approaches in its specific implementation mechanism, rather than merely adopting foreground-background modeling.

  • Furthermore, BIS differs from both PFNet and conventional feature reweighting methods. PFNet suppresses interference through dual-branch foreground-background context exploration, whereas BIS generates a single gating map based on CDEM outputs and suppresses background interference through residual subtraction while preserving the original feature pathway. Unlike conventional reweighting methods that directly modulate original features, BIS also employs lightweight local refinement to integrate useful responses. We have clarified these differences in Section 3.2.2 of the revised manuscript to further explain its design motivation and mechanism.

In summary, our method does not simply combine existing prototype modeling, foreground-background exploration, and prediction-guided mechanisms. Instead, it is designed to address the needs of lightweight COD by introducing distinct prediction-guided cross-layer interaction and background interference suppression mechanisms for discriminative evidence mining.

(2) Response to the Contributions of Individual BIS Components

Thank you for your comments on the BIS design. To investigate the impact of BIS’s internal mechanisms on model performance, we conduct two additional ablation experiments by removing the gating operation (w/o gate) and the refinement operation (w/o refine), respectively. The experimental results are shown in the table below.

SettingsCAMOCHAMELEONCOD10KNC4K
Sm↑S_m\uparrowFβw↑F_\beta^w\uparrowM↓M\downarrowEm↑E_m\uparrowSm↑S_m\uparrowFβw↑F_\beta^w\uparrowM↓M\downarrowEm↑E_m\uparrowSm↑S_m\uparrowFβw↑F_\beta^w\uparrowM↓M\downarrowEm↑E_m\uparrowSm↑S_m\uparrowFβw↑F_\beta^w\uparrowM↓M\downarrowEm↑E_m\uparrow
w/o gate0.8750.8320.0480.9240.9070.8610.0230.9530.8710.7920.0220.9340.8920.8490.0310.937
w/o refine0.8770.8370.0470.9290.9170.8740.0200.9620.8730.7940.0220.9350.8920.8460.0310.937
DMRNet0.8840.8460.0440.9330.9180.8750.0190.9620.8730.7960.0220.9360.8930.8490.0310.938

Both ablation settings lead to performance degradation, with removing the gate operation causing a greater overall decline than removing the refinement operation. The gate operation generates a gating map from CDEM outputs to suppress potential background interference while preserving the original feature pathway, thereby improving the reliability of discriminative features. This demonstrates that the gating operation plays a more important role in BIS, while the refinement operation provides additional performance gains.

In addition, we conduct qualitative visual comparisons of the two ablation settings. As shown in Fig. 4, which has been newly added to the revised manuscript, removing either the gate or refine operation introduces varying degrees of background noise into the predicted masks, whereas the complete BIS produces more precise predictions across different camouflage scenarios. This further qualitatively demonstrates the contributions of both operations to background interference suppression.

Based on these results, we have added the internal ablation experiments of BIS, the corresponding analysis, and qualitative visualizations to Section 4.3 of the revised manuscript to further validate the rationale and effectiveness of the internal design of BIS.

AuthorsOpenReview ↗

Response to Weakness (2) from Reviewer v72E

Response to the Contributions of the Key MFC Components

Thank you for pointing out the limitations of the original ablation study. To verify the effectiveness of the internal mechanisms of MFC, we conduct ablation experiments by individually removing foreground-background prototype competition (w/o F/B C), confidence weighting (w/o CW), and uncertainty-guided weighting (w/o UG). The experimental results are presented in the table below.

SettingsCAMOCHAMELEONCOD10KNC4K
Sm↑S_m\uparrowFβw↑F_\beta^w\uparrowM↓M\downarrowEm↑E_m\uparrowSm↑S_m\uparrowFβw↑F_\beta^w\uparrowM↓M\downarrowEm↑E_m\uparrowSm↑S_m\uparrowFβw↑F_\beta^w\uparrowM↓M\downarrowEm↑E_m\uparrowSm↑S_m\uparrowFβw↑F_\beta^w\uparrowM↓M\downarrowEm↑E_m\uparrow
w/o F/B C0.8790.8380.0460.9310.9160.8700.0210.9610.8730.7940.0220.9350.8930.8480.0310.938
w/o CW0.8780.8380.0460.9290.9130.8690.0210.9600.8720.7950.0220.9350.8920.8470.0310.937
w/o UG0.8790.8410.0460.9310.9150.8710.0210.9600.8710.7930.0230.9330.8920.8490.0320.936
DMRNet0.8840.8460.0440.9330.9180.8750.0190.9620.8730.7960.0220.9360.8930.8490.0310.938

As shown in the table, removing any of the three mechanisms leads to performance degradation, confirming their contributions to MFC. Among the two mechanisms you specifically highlighted, confidence-weighted prototype extraction contributes more to overall performance than foreground-background prototype competition. This is because confidence weighting improves the reliability of foreground and background prototypes, providing a more reliable basis for subsequent prototype competition and discriminative evidence mining. In comparison, prototype competition exploits the constructed prototypes to provide discriminative cues, while uncertainty-guided weighting adjusts spatial routing weights according to prediction uncertainty. Overall, the three mechanisms play complementary roles in prototype reliability, foreground-background discrimination, and feature routing, respectively, and jointly enhance discriminative evidence mining.

Following your suggestion, we have added the above three ablation experiments, complete experimental results, and corresponding analyses to the Appendix 10 to further validate the individual contributions and collaborative effects of the MFC components.

AuthorsOpenReview ↗

Request Feedback from Reviewer v72E

Dear Reviewer v72E, 

Thank you again for your thoughtful comments on our submission. As the discussion period is coming to a close, we would like to ask whether our responses have addressed your concerns. If any points remain unclear or if you have additional questions, we would greatly appreciate your guidance and will be happy to clarify further. If you feel that our revisions and explanations have resolved the issues you raised, we would be grateful if you could consider updating your evaluation accordingly. 

We sincerely look forward to your feedback, and thank you for your time and effort throughout this process. 

Best regards, 

Authors of submission36108

Comments on the paper

2 comments
AuthorsOpenReview ↗

Response to All Reviewers

Dear Reviewers,

First, we sincerely thank all reviewers for carefully evaluating our work and providing valuable comments. These comments have helped us further examine the limitations of our method design, experimental validation, and manuscript presentation, and have provided important guidance for improving our work. We have carefully considered all the comments.

This work addresses a key challenge in lightweight camouflaged object detection, i.e., how to effectively mine and preserve subtle discriminative information between camouflaged objects and complex backgrounds under limited model parameters and computational budgets. To address this issue, we propose a Discriminative Evidence Mining and Reconstruction Network (DMRNet). Its core idea is to mine discriminative information through prediction-guided cross-layer foreground-background prototype competition, and further enhance feature representations through background interference suppression and evidence-guided hierarchical feature reconstruction.

We would like to once again express our sincere gratitude to all the reviewers for taking the time to review our work and provide constructive comments.

AuthorsOpenReview ↗

Summary of Revisions

Following the reviewers' comments and suggestions, we have carefully revised and supplemented the manuscript. The main revisions are summarized as follows.

    1. Revision of the Overall Architecture Figure. We have redrawn Fig. 2 to provide a clearer illustration of the overall network architecture, making the connections and information flow among different components easier to understand.
    1. Clarification of the Novelty of BIS. We have added a description in Section 3.2.2 to clarify the differences between BIS and existing background suppression and feature reweighting methods, further explaining its design motivation and operational mechanism.
    1. Description of Checkpoint Selection. We have added details in Section 4.1 regarding the checkpoint selection criterion and its use, clarifying the evaluation procedure during training and testing.
    1. Ablation Study and Visualization of BIS Components. We have added ablation experiments on the internal components of BIS in Section 4.3, analyzed their respective contributions, and included corresponding qualitative visualizations to further demonstrate their effectiveness.
    1. Sensitivity Analysis of the Temperature Coefficient. We have added a dataset-wise table in Appendix 5 reporting the detailed performance variations under different temperature coefficients across the four benchmark datasets, together with the corresponding analysis.
    1. Analysis of Clipping Bounds. We have added a new Appendix 6 to investigate the effects of different clipping-bound settings on model performance and further justify the adopted parameter choices.
    1. Comparison of CDEM with Conventional Attention Modules. We have added Appendix 9, where CDEM is compared with conventional attention modules, including CBAM and SENet, through replacement experiments to further demonstrate the characteristics and effectiveness of our design.
    1. Ablation Study of MFC Components. We have added Appendix 10, which presents ablation experiments on the individual mechanisms within MFC to investigate their respective contributions to overall model performance.
    1. Analysis of the Design Choices in Eqs. (7)–(9). We have added Appendix 11 to explain the motivations behind the hyperbolic tangent function, quadratic uncertainty measure, linear combination, and clipping operation. Additional experiments using alternative formulations are conducted to validate the effectiveness of these design choices.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.