acceptodds
Under review as a conference paper at ICLR 2027

Non-Equilibrium Diffusion Pseudo-labelling and Relative Preference Optimisation for Unsupervised Camouflaged Object Detection

Abstract

Unsupervised Camouflaged Object Detection (UCOD) is challenged by extreme visual entanglement between objects and surroundings. Existing methods struggle to integrate the heterogeneous cues of intra-object multi-camouflage, yielding fragmented pseudo-labels plagued by structural omissions and foreground-background confusion, which ultimately compels models to passively overfit initial errors rather than recover the complete object structure. To address these problems, we propose DES-UCOD, a novel Discover-Explore-Specialise framework for Unsupervised Camouflaged Object Detection that shifts from passive label fitting to active object discovery. Firstly, to resolve feature interference under multi-camouflage conditions and provide comprehensive candidate supervision, we propose the Non-Equilibrium Diffusion Pseudo-labelling (NEDP) strategy, which captures transient structural dynamics prior to global equilibrium to theoretically unify fragmented camouflage cues and discover robust structural anchors to generate high-quality pseudo-labels. Secondly, to mitigate the negative impact of pseudo-label noise and significantly alleviate confirmation bias, we propose the Stochastic Relative Preference Optimisation (S-RPO) strategy, an active reward-driven sampling mechanism that compels the model to explore decision boundaries highly robust against diverse visual variations during training. Finally, to enhance the model's capacity for handling diverse camouflage characteristics, we propose a Spatial-Cue Mixture-of-Experts (SC-MoE) decoder, which specialises in routing spatially heterogeneous features at each location, ensuring that distinct experts fulfil their respective roles to generate accurate predictions. Extensive experiments demonstrate that our model consistently outperforms various UCOD methods across multiple datasets to achieve state-of-the-art performance.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Unsupervised Camouflaged Object Detection

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_Xp3sOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes DES-UCOD for unsupervised camouflaged object detection, combining pseudo-label generation, preference optimisation, and a mixture-of-experts decoder, with experiments on four datasets.

Critical strengths

The method achieves promising results on multiple datasets and includes ablation studies to support its effectiveness

Critical weaknesses
  1. Unclear experimental fairness. The paper does not clearly explain whether the compared methods use consistent experimental settings, which may affect the reliability of the comparisons.
  2. Unclear hyperparameter selection. Table 7 reports hyperparameter experiments on COD10K, but it is unclear whether an independent validation set was used to select the final parameters.
Questions for authors

Q1. Were the baseline methods evaluated under consistent experimental settings, or were their results taken from the original papers? Q2. Were the final hyperparameters selected using an independent validation set rather than the test set?

2 comments
AuthorsOpenReview ↗

Author Responses

Q1. Unclear experimental fairness and consistency of experimental settings.

We sincerely thank the reviewer for raising this important concern. We would like to clarify that our experiments strictly follow the experimental protocol of DualUCOD, ensuring consistency across key experimental settings.

Firstly, we maintain consistent experimental configurations, including but not limited to input image resolution (476 × 476), dataset configurations, training and testing splits, data preprocessing, and evaluation protocols. These settings help minimize the influence of unrelated experimental factors and ensure a fair assessment of our proposed method.

Secondly, for competing methods, we prioritize the results officially reported in their original publications when the experimental settings and evaluation protocols are consistent with ours. When discrepancies exist in experimental configurations, we re-run the corresponding methods under unified experimental settings, including a standardized input resolution of 476 × 476, to ensure fair and controlled comparisons. All methods are evaluated on the same benchmark datasets using consistent evaluation metrics. Furthermore, as shown in Table 1, we separately organize comparisons based on DINO and DINOv2 backbones to account for architectural differences and enable more meaningful comparisons.

Overall, our experiments follow consistent benchmark and evaluation protocols. By prioritizing officially published results under comparable settings and conducting controlled re-evaluations when necessary, we aim to ensure the fairness, consistency, and reliability of our experimental comparisons.

We sincerely appreciate the reviewer's constructive suggestion.

Q2. Unclear hyperparameter selection and whether an independent validation set was used.

We sincerely thank the reviewer for this valuable comment. We would like to clarify that an independent validation subset was randomly selected from the COD10K-Train and CAMO-Train datasets for hyperparameter tuning, following a validation strategy similar to previous work [1].

Firstly, all hyperparameters reported in Table 7 were empirically determined based on this validation subset to balance performance and computational efficiency.

Secondly, once the hyperparameters were selected, the same configuration was consistently applied across all experiments, ensuring fairness, consistency, and reproducibility.

Overall, our hyperparameter selection follows an independent validation-subset protocol. We acknowledge that this procedure was not sufficiently clarified in the original manuscript and will explicitly describe the validation strategy in the revised version to improve transparency and reproducibility.

We sincerely appreciate the reviewer's constructive suggestion.

Reference

[1] Y. Yang, J. Tian, S. Zhang, and L. Cao, "Discover, Segment, and Select: A Progressive Mechanism for Zero-shot Camouflaged Object Segmentation," Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 34745–34754, 2026.

Dear Reviewer, As the discussion period is drawing to a close, we would like to gently follow up to see if our previous responses and additional experiments have addressed your concerns.

We are more than happy to provide any further clarifications if you have any remaining questions. Thank you again for your time and constructive feedback.

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

This paper addresses the challenges in unsupervised camouflaged object detection (UCOD), specifically the fragmentation of structural cues, i.e., arising from varying camouflage characteristics across different parts of an object, incomplete pseudo-labels, and learning biases induced by noisy supervision. To tackle these, the authors propose the DES-UCOD framework, which employs unbalanced diffusion-based pseudo-label generation, stochastic relative preference optimization (S-RPO), and a spatially-conditioned mixture-of-experts (SC-MoE) decoder. These components are designed to improve structural discovery, robust learning under noisy supervision, and the modeling of spatially heterogeneous features, respectively. The research addresses a significant problem: how to recover complete objects from local cues that are highly confounded with the background, in the absence of human annotations.

Critical strengths

The paper’s primary strengths lie in its clear research motivation, the strong alignment between the proposed methods and the identified challenges, and a comprehensive set of benchmark comparisons and ablation studies. Within the DINOv2 setting, the method outperforms existing approaches across most evaluation metrics on four datasets, demonstrating robust overall effectiveness. Notably, the appendix includes an evaluation of pseudo-label quality and an analysis of the relationship between foreground spectral energy coverage, spectral dispersion, and the benefits of multi-modal aggregation. This provides targeted evidence, i.e., beyond mere performance gains, for the core premise that foreground information is dispersed across multiple spectral modes. Furthermore, the paper acknowledges that performance may decline for samples with low spectral dispersion, thereby clarifying the conditional nature of the method's effectiveness.

Critical weaknesses

A limitation of the study is that the experimental evidence supporting the overall framework is stronger than the validation of individual innovative mechanisms. The unbalanced diffusion process utilizes a mathematical formulation linked to the diagonal of the heat kernel; however, the absence of an ablation that isolates the diffusion-weighting component makes it difficult to discern whether performance gains stem from multi-spectral modal aggregation or the finite-time diffusion process itself. Similarly, the S-RPO component lacks internal ablation studies or controlled experiments regarding pseudo-label noise, while the SC-MoE component lacks comparisons against decoders with equivalent parameter counts or baselines that use fixed fusion rather than spatial routing. Consequently, the claims regarding the mitigation of confirmation bias and expert specialization remain insufficiently substantiated. It is recommended that the authors include these additional controls, analyze the evolution of errors during the training process, provide visualizations of the expert routing, and report the mean and variance across multiple experimental runs. While existing evaluation metrics are suitable for assessing overall segmentation quality, the evaluation framework should be expanded to include metrics for target omission (missed detections), boundary quality, and performance stratified by the level of camouflage difficulty. Furthermore, the current "spectral dispersion" metric does not directly capture the proportion of a target's area that is visually similar to the background. Additionally, the performance advantage observed under the DINO setting is not entirely stable; the authors should further clarify the method's reliance on pre-trained representations and explicitly define the boundaries of its substantive innovation relative to existing techniques.

Questions for authors

The authors are requested to clarify the following points: (1) When keeping graph construction, the number of spectral modes, and post-processing consistent, what is the independent gain provided by finite-time diffusion compared to diffusion-free weighted aggregation (i.e., at t=0t=0)? Do the actual data satisfy the effective time-window requirements stipulated by the theoretical analysis? (2) How much of the performance gain in S-RPO is attributable to each component, i.e., multi-view augmentation, confidence weighting, random sampling, and relative preference learning? Furthermore, how does the mechanism prevent the reinforcement of incorrect pseudo-labels when multiple views consistently support them? (3) Does SC-MoE offer a clear advantage over a multi-branch decoder with an equivalent number of parameters or a fixed fusion strategy? Is there direct evidence that the individual experts have developed the intended spatial specialization? (4) Is it possible to define a reproducible criterion for target-background similarity? Could the authors report the proportion of similar regions (or other direct statistics on camouflage difficulty) for each dataset, and examine whether performance gains vary according to this difficulty? (5) Are the primary results stable across different random seeds, and are the observed performance margins, where small, statistically significant? Were hyperparameters determined using an independent validation set? Finally, are the backbone architectures, input resolutions, and training configurations of the compared methods sufficiently comparable?

5 comments
AuthorsOpenReview ↗

Author Responses

Q1. Independent contribution of finite-time diffusion and empirical validation of the effective diffusion regime.

We sincerely thank the reviewer for this valuable suggestion. To isolate the contribution of finite-time diffusion, we compare t=0t=0 and t=10t=10 while keeping the graph construction, number of retained spectral modes (m=10m=10), and Structural-to-Semantic Calibration (SSC) post-processing consistent.

1. Independent Contribution of Finite-Time Diffusion.

Table R1. Downstream segmentation performance under different diffusion settings.

DatasetttM↓M\downarrowFβw↑F_\beta^w\uparrowEϕ↑E_\phi\uparrowSα↑S_\alpha\uparrow
COD10K00.0430.6630.8670.772
100.0350.6960.8870.793
NC4K00.0680.7160.8710.780
100.0470.7830.9110.827

As shown in Table R1, finite-time diffusion consistently improves all four metrics. Specifically, FβwF_\beta^w increases by 0.033 on COD10K and 0.067 on NC4K, while MAE decreases by 0.008 and 0.021, respectively. These results demonstrate the independent empirical benefits of finite-time spectral weighting beyond diffusion-free multi-mode aggregation.

2. Empirical Analysis of Finite-Time Spectral Weighting.

We further conducted a spectral analysis on 4,026 valid training images using identical graph construction, spectral decomposition, and post-processing.

Table R2. Empirical analysis of finite-time spectral weighting.

ttR11,2R_{11,2} (Median)NeffN_{\mathrm{eff}} (Mean)Fβw↑F_\beta^w\uparrowSα↑S_\alpha\uparrow
01.00010.0000.5670.698
100.0162.6090.5900.718

Lower R11,2R_{11,2} indicates stronger relative attenuation of higher-frequency modes, while lower NeffN_{\mathrm{eff}} indicates more concentrated spectral weighting. Neither quantity is inherently better.

As shown in Table R2, finite-time diffusion enhances relative low-frequency selectivity while retaining contributions from multiple spectral modes. Meanwhile, pseudo-label quality improves by 0.0229 in FβwF_\beta^w (95% CI: [0.0150, 0.0312]) and 0.0194 in SαS_\alpha (95% CI: [0.0143, 0.0246]), based on 10,000 paired bootstrap resamples.

3. Empirical Validation and Theoretical Interpretation.

Our additional experiments provide strong empirical evidence supporting the effectiveness of finite-time diffusion. The observed spectral reweighting, improved pseudo-label quality, and consistent downstream performance gains demonstrate that t=10t=10 is an effective operating point for NEDP.

Furthermore, 48.68% of valid images satisfy the joint spectral coverage and dispersion criteria, while 61.40% satisfy the predefined spectral-separation diagnostic. These findings support the practical relevance of the spectral assumptions and the proposed mechanism of balancing spectral selectivity with complementary structural information.

Based on our empirical evaluations, t=10t=10 was identified as an effective operating point for finite-time diffusion, yielding consistent improvements in pseudo-label quality and downstream segmentation performance. These findings provide strong empirical support for the effectiveness of the selected diffusion time and the proposed non-equilibrium diffusion mechanism.

Overall, these additional experiments substantiate the independent contribution of finite-time spectral weighting and provide empirical support for the spectral-selectivity mechanism underlying NEDP.

We sincerely appreciate the reviewer's constructive suggestion.

AuthorsOpenReview ↗

Author Responses (continued)

Q2. Component-wise contributions of S-RPO and robustness to incorrect pseudo-labels.

We sincerely thank the reviewer for this valuable suggestion. We conducted additional component-wise ablations on COD10K to evaluate the individual contributions of S-RPO.

1. Component-wise Contributions.

Table R3. Component-wise ablation study of S-RPO on COD10K.

ConfigurationM↓M\downarrowFβw↑F_\beta^w\uparrowEϕ↑E_\phi\uparrowSα↑S_\alpha\uparrow
w/o S-RPO0.0400.6770.8770.780
w/o Multi-view Augmentation0.0370.6810.8800.785
w/o Confidence Weighting0.0390.6820.8700.787
w/o Stochastic Sampling0.0380.6790.8670.786
w/o Relative Preference0.0390.6820.8730.787
Full Model0.0350.6960.8870.793

As shown in Table R3, all four components contribute positively. Removing multi-view augmentation, confidence weighting, stochastic sampling, and relative preference learning decreases FβwF_\beta^w by 0.015, 0.014, 0.017, and 0.014, respectively, confirming their complementary roles.

2. Robustness to Incorrect Pseudo-labels.

Assessing pseudo-label correctness without ground-truth annotations remains a fundamental challenge in unsupervised learning. S-RPO partially addresses this issue through three mechanisms: Eq. (15) estimates cross-view predictive support with leave-one-view-out and stop-gradient operations; Eq. (16) reduces the reward contribution of poorly supported regions; and Eqs. (14), (17), and (18) introduce stochastic exploration and relative preference optimization to alleviate confirmation bias.

Nevertheless, cross-view consistency does not guarantee semantic correctness. Consistently incorrect pseudo-labels may still receive high confidence. Thus, our method mitigates view-inconsistent errors rather than completely eliminating systematic errors. We will further explore uncertainty-aware confidence calibration and adaptive pseudo-label correction to better identify systematic errors while maintaining annotation-free learning.

We sincerely appreciate the reviewer's constructive suggestion.

Q3. Effectiveness and spatial specialization of SC-MoE.

We sincerely thank the reviewer for this constructive suggestion. We conducted additional comparisons with a multi-branch decoder and a fixed equal-weight fusion strategy on COD10K.

Table R4. Comparison of decoder architectures and fusion strategies on COD10K.

ConfigurationM↓M\downarrowFβw↑F_\beta^w\uparrowEϕ↑E_\phi\uparrowSα↑S_\alpha\uparrow
Multi-branch Decoder0.04190.66920.85960.7784
Fixed Equal-weight Fusion0.04140.66960.86010.7784
SC-MoE (Ours)0.04070.67080.86140.7792

As shown in Table R4, SC-MoE achieves better performance across all four metrics, supporting the effectiveness of spatially adaptive fusion over conventional multi-branch processing and fixed fusion.

Regarding spatial specialization, Figure 3 illustrates the distinct feature-processing mechanisms of the three experts. Specifically, LCE captures local contrast, RSE models regional structural information, and GCE incorporates global contextual cues, as formulated in Eqs. (19)–(21). These complementary representations are adaptively integrated through the spatial routing mechanism in Eq. (22).

Furthermore, the expert ablations in the original Table 4 show that removing LCE, RSE, or GCE reduces FβwF_\beta^w from 0.696 to 0.680, 0.670, and 0.671, respectively. Together with the architectural illustration in Figure 3, these results support the complementary roles of the three experts and the benefits of spatially adaptive fusion.

We sincerely appreciate the reviewer's valuable suggestion.

AuthorsOpenReview ↗

Author Responses (continued)

Q4. Quantitative characterization of target–background similarity and difficulty-dependent performance.

We sincerely thank the reviewer for this valuable suggestion. To quantitatively evaluate camouflage difficulty, we conducted additional analyses on CAMO-Train and COD10K-Train (4,040 images) without retraining the model.

1. Reproducible Target–Background Similarity.

We define a feature-based similarity criterion using frozen DINOv2 representations. Foreground and background patches are identified using GT masks with a 75% coverage threshold, solely for post-hoc analysis. For each foreground patch, we calculate the average cosine similarity with its five spatially nearest background patches:

\frac{f_u^{\top}f_v}{\\|f_u\\|_2\\|f_v\\|_2}$$ The image-level similarity is defined as: $$H(I) = \frac{1}{|F_I|}\sum_{u \in F_I}h(u)$$ The proportion of high-similarity foreground patches is defined as: $$R_{\tau}(I) = \frac{1}{|F_I|} \sum_{u \in F_I}\mathbf{1}[h(u)\geq\tau]$$ Here, the shared threshold $\tau$ corresponds to the 75th percentile of foreground-patch similarities in the combined training set. **Table R6. Target–background similarity statistics on training datasets.** | Dataset | Valid / Total | Mean Similarity $H$ | High-Similarity Patches (%) | | :--- | :---: | :---: | :---: | | CAMO-Train | 1,000 / 1,000 | 0.2961 | 26.97 | | COD10K-Train | 2,975 / 3,040 | 0.3035 | 23.67 | | **Combined** | **3,975 / 4,040** | **0.3016** | **25.00** | As shown in Table R6, both datasets exhibit substantial target–background similarity. The 65 excluded images lack valid foreground patches under the predefined criterion. The combined 25% proportion reflects the percentile-based definition rather than an absolute camouflage threshold. **2. Performance Gains Across Camouflage Difficulty Levels.** We divided the **3,975 valid images** into three equally sized groups according to $H(I)$ and compared NEDP with the conventional **Fiedler spectral partition**, using the same frozen features and graph eigendecomposition. **Table R7. Difficulty-dependent pseudo-label improvements over the Fiedler baseline.** | Similarity | Images | Fiedler $F_\beta^w\uparrow$ | NEDP $F_\beta^w\uparrow$ | $\Delta F_\beta^w$ [95% CI] | $\Delta S_\alpha\uparrow$ | | :--- | :---: | :---: | :---: | :---: | :---: | | Low | 1,325 | 0.7014 | 0.7402 | +0.0388 [0.0258, 0.0523] | +0.0199 | | Medium | 1,325 | 0.5500 | 0.6464 | +0.0964 [0.0828, 0.1103] | +0.0686 | | High | 1,325 | 0.2633 | 0.4035 | **+0.1402 [0.1279, 0.1523]** | **+0.1306** | *Confidence intervals are obtained using 10,000 image-level paired bootstrap resamples.* As shown in Table R7, **NEDP's advantage increases with camouflage difficulty**. Specifically, the $F_\beta^w$ improvement increases from **0.0388 to 0.1402**, while the $S_\alpha$ improvement increases from **0.0199 to 0.1306** from low- to high-similarity regions. These findings support our motivation for **multi-mode structural discovery**, demonstrating that NEDP provides greater relative benefits when foreground and background features become increasingly similar, although absolute performance remains lower in more challenging cases. **Overall,** the additional experiments establish a reproducible similarity criterion and provide empirical evidence that NEDP's improvements are particularly pronounced under challenging camouflage conditions. These results also complement the spectral-dispersion analysis in **Appendix A.2.3**. GT masks are used exclusively for post-hoc analysis and evaluation, without participating in pseudo-label generation or model optimisation. We sincerely appreciate the reviewer's constructive suggestion.
Reviewer_T8gCOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes DES-UCOD for unsupervised camouflaged object detection. The framework consists of three core components: Non-Equilibrium Diffusion Pseudo-labelling (NEDP) to aggregate low-frequency graph modes via finite-time diffusion, Stochastic Relative Preference Optimisation (S-RPO) to alleviate confirmation bias across aligned views, and a Spatially Conditioned Mixture of Experts (SC-MoE) decoder. The authors provide oracle spectral derivations and report consistent improvements across four benchmarks.

Critical strengths

1.The framework achieves solid and consistent gains across four COD benchmarks (e.g., F_beta^w improves from 0.635 to 0.696 on COD10K). In addition, Table 9 directly validates that NEDP improves initial pseudo-label quality prior to model training.

2.Appendix A.2.3 provides rigorous diagnostics across 4,026 training images, showing that multi-mode aggregation significantly improves coverage and quality for high-dispersion cases while transparently reporting performance drops on low-dispersion instances.

Critical weaknesses

1.The structural response st(u)=∑k=2m+1e−tλkϕk(u)2s_t(u) = \sum_{k=2}^{m+1} e^{-t\lambda_k} \phi_k(u)^2 lacks an unweighted baseline (t=0t=0) under identical graph and SSC post-processing settings. For a normalized graph Laplacian, eigenvalues satisfy λk∈[0,2]\lambda_k \in [0, 2]. With t=10t=10, modes with λk≥0.5\lambda_k \ge 0.5 are suppressed by e−tλk≤0.0067e^{-t\lambda_k} \le 0.0067. It is thus difficult to determine whether improvements stem from non-equilibrium diffusion dynamics or simply aggressive mode truncation equivalent to keeping only 2–3 modes.

2.Cross-view support derives from the same network checkpoint and shares the same static pseudo-labels. Furthermore, clipping the continuous Binary Concrete log-density to [−5,0][-5, 0] in Eq. (18) is mathematically problematic: continuous densities can strictly exceed 1 (e.g., log⁡p≈+1.433\log p \approx +1.433 when P=0.5,τs=0.5,M=0.01P=0.5, \tau_s=0.5, M=0.01). Capping positive values to 0 zeros out gradients (∂clip/∂θ=0\partial \text{clip}/\partial \theta = 0) on confident predictions, distorting the intended optimization behavior.

Questions for authors
  1. Under identical graph construction, m=10m=10, and SSC post-processing, what are the pseudo-label quality metrics and downstream detection performance when setting t=0t=0 (unweighted sum of 10 modes) versus t=10t=10?

  2. Does the actual backpropagation pass through clip(log⁡p,−5,0)\text{clip}(\log p, -5, 0)? What percentage of training samples fall above 0 (receiving zero gradient), and how does the model perform without the upper-bound clipping?

3 comments
AuthorsOpenReview ↗

Author Responses

Q1. Independent contribution of finite-time diffusion to pseudo-label quality and downstream performance.

We sincerely thank the reviewer for this constructive suggestion. To isolate the contribution of finite-time diffusion, we compare t=0t=0 and t=10t=10 while keeping the graph construction, number of retained spectral modes (m=10m=10), and SSC post-processing identical.

1. Comparison of Pseudo-label Quality.

According to Eq. (11), the structural response is defined as

st(u)=∑k=2m+1e−tλkϕk(u)2.s_t(u)=\sum_{k=2}^{m+1}e^{-t\lambda_k}\phi_k(u)^2.

At t=0t=0, all ten modes are equally weighted, whereas t=10t=10 introduces frequency-dependent attenuation to emphasize lower-frequency structural information.

Table R1. Comparison of pseudo-label quality on CAMO-Train and COD10K-Train (4,040 images).

ttSα↑S_\alpha\uparrowFβw↑F_\beta^w\uparrowFβm↑F_\beta^m\uparrowFβa↑F_\beta^a\uparrowFβmax⁡↑F_\beta^{\max}\uparrow
00.6980.5650.6200.6220.622
100.7170.5880.6270.6290.629

As shown in Table R1, finite-time diffusion improves all five reported metrics, particularly FβwF_\beta^w (+0.023) and SαS_\alpha (+0.019). A paired bootstrap analysis on 4,026 valid images (10,000 resamples) further confirms positive gains in FβwF_\beta^w (+0.0229, 95% CI: [0.0150, 0.0312]) and SαS_\alpha (+0.0194, 95% CI: [0.0143, 0.0246]). Other pseudo-label metrics exhibit some trade-offs.

2. Comparison of Downstream Segmentation Performance.

We further evaluate whether finite-time spectral weighting improves downstream segmentation under consistent network and training configurations.

Table R2. Downstream segmentation performance under different diffusion settings.

DatasetttM↓M\downarrowFβw↑F_\beta^w\uparrowEϕ↑E_\phi\uparrowSα↑S_\alpha\uparrow
CHAMELEON00.0560.7400.9010.806
100.0360.7960.9270.851
CAMO00.1040.6600.8310.730
100.0810.7230.8550.774
COD10K00.0430.6630.8670.772
100.0350.6960.8870.793
NC4K00.0680.7160.8710.780
100.0470.7830.9110.827

As shown in Table R2, t=10t=10 consistently outperforms t=0t=0 across all four benchmarks. Notably, FβwF_\beta^w improves by 0.033 on COD10K and 0.067 on NC4K, while MAE decreases by 0.008 and 0.021, respectively.

Overall, these controlled experiments demonstrate that finite-time spectral weighting provides additional benefits beyond multi-mode aggregation alone, improving both structural pseudo-label quality and downstream segmentation performance.

We sincerely appreciate the reviewer's valuable suggestion, which has helped us further clarify the independent contribution of finite-time diffusion.

AuthorsOpenReview ↗

Author Responses (continued)

Q2. Gradient propagation through log-density clipping and the effect of upper-bound clipping.

We sincerely thank the reviewer for this insightful question. To clarify the role of log-density clipping in S-RPO, we examined the actual gradient behavior and conducted an additional ablation experiment on COD10K.

1. Gradient Propagation and Clipping Statistics.

The clipping operation clip⁡(log⁡p,−5,0)\operatorname{clip}(\log p,-5,0) is included in the actual backpropagation path. Values outside [−5,0][-5,0] receive zero gradient through the corresponding clipped log-density term.

We analyzed a large number of training batches, covering approximately 10.98 billion batch-view-pixel elements. The results show that 93.99% of raw log-density values exceed 0, 0.06% fall below −5, and 5.95% remain within the unclipped interval.

Importantly, a high clipping ratio does not necessarily indicate ineffective optimization. Since the Relaxed Bernoulli distribution is continuous, its probability density can exceed 1, resulting in positive log-density values. Although upper-bound clipping suppresses a substantial proportion of direct log-density gradients, the remaining unclipped elements continue to contribute to preference optimization, while structural supervision and entropy regularization provide complementary learning signals.

2. Effect of Upper-bound Clipping.

To further investigate its influence, we evaluated a variant that removes only the upper clipping bound while retaining the lower bound at −5.

Table R. Effect of upper-bound clipping on COD10K.

ConfigurationM↓M\downarrowFβw↑F_\beta^w\uparrowEϕ↑E_\phi\uparrowSα↑S_\alpha\uparrow
Full Model (Ours)0.0350.6960.8870.793
w/o Upper-bound Clipping0.0880.1110.3240.477

The variant without upper-bound clipping exhibits substantially degraded segmentation performance compared with the reported full-model results. Furthermore, our controlled gradient analysis shows that removing the upper bound increases the mean S-RPO gradient norm from 0.2293 to 13.5501, approximately a 59-fold increase, without producing NaN or Inf values.

These observations support the importance of upper-bound clipping in controlling excessive gradient contributions and maintaining stable preference-guided optimization.

3. Limitations and Future Work.

We also recognize that the relatively high clipping ratio indicates an opportunity to further improve the utilization of preference-learning signals. Although the current clipping strategy effectively controls gradient magnitudes, it inevitably suppresses direct gradient contributions from many high-density elements.

In future work, we plan to investigate adaptive clipping thresholds and smooth gradient-bounding strategies that can preserve more informative preference gradients while preventing excessively large updates. We will also explore temperature-aware log-density calibration to better balance gradient utilization and optimization stability. These improvements may further enhance the efficiency and robustness of S-RPO without compromising its current advantages.

Overall, our additional analysis demonstrates the practical importance of upper-bound clipping while identifying a promising direction for further improving gradient utilization in preference-guided optimization.

We sincerely appreciate the reviewer's constructive suggestion, which has helped us clarify both the optimization behavior and potential future improvements of S-RPO.

Dear Reviewer, As the discussion period is drawing to a close, we would like to gently follow up to see if our previous responses and additional experiments have addressed your concerns.

We are more than happy to provide any further clarifications if you have any remaining questions. Thank you again for your time and constructive feedback.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

General Response

We sincerely thank reviewers T8gC, fzgt, and Xp3s for their constructive feedback and for recognizing our benchmark improvements, spectral diagnostics (T8gC), clear motivation (fzgt), and promising ablation results (Xp3s). Your suggestions have significantly strengthened our experimental evidence and methodological clarity.

Below is a summary of our main updates:

  • Isolated the contribution of finite-time diffusion (T8gC, fzgt): Comparing t=10t=10 to t=0t=0 (keeping graph construction, m=10m=10 modes, and SSC post-processing identical), finite-time weighting improved pseudo-label FβwF_\beta^w from 0.565 to 0.588 and SαS_\alpha from 0.698 to 0.717. This also drove downstream gains across all four benchmarks (e.g., FβwF_\beta^w +0.033 on COD10K and +0.067 on NC4K). Spectral diagnostics and paired-bootstrap analysis on 4,026 training images further validate the benefits of spectral weighting over equal-weight multi-mode aggregation.

  • Examined gradient behavior and upper-bound clipping (T8gC): We verified that log-density clipping participates in backpropagation, with 93.99% of pixel-level raw log-density values exceeding zero, explicitly suppressing direct gradients. Removing the upper clipping bound (while keeping the lower bound at −5) spiked the mean S-RPO gradient norm from 0.2293 to 13.5501 (~59-fold) and collapsed COD10K FβwF_\beta^w from 0.696 to 0.111. This confirms the upper bound's critical role in controlling gradient magnitudes, while acknowledging current limitations in gradient utilization.

  • Added component-wise S-RPO ablations and clarified robustness (fzgt, T8gC): Removing multi-view augmentation, confidence weighting, stochastic sampling, or relative preference learning reduced COD10K FβwF_\beta^w from 0.696 to 0.681, 0.682, 0.679, and 0.682, respectively. We also clarified how leave-one-view-out support estimation, stop-gradients, and confidence-weighted rewards effectively mitigate view-inconsistent errors.

  • Strengthened decoder comparisons and adaptive expert fusion (fzgt): Compared to a multi-branch decoder and fixed equal-weight fusion, SC-MoE achieved higher performance (Fβw=0.6708F_\beta^w=0.6708 vs. 0.6692 and 0.6696). Alongside our expert-removal ablations, these results further validate the benefits of complementary expert processing (local-contrast, regional-structure, global-context) and spatially adaptive fusion.

  • Introduced reproducible camouflage-difficulty analysis (fzgt): Using frozen DINOv2 features, we measured target–background similarity between foreground patches and their nearest background neighbors. High-similarity patches comprised 26.97% of CAMO-Train and 23.67% of COD10K-Train. Dividing 3,975 training images into three similarity groups, NEDP’s pseudo-label FβwF_\beta^w improvements over the Fiedler baseline scaled from +0.0388 (low similarity) to +0.0964 (medium) and +0.1402 (high). This confirms our method yields larger gains on harder camouflage. (GT masks were used solely for this post-hoc evaluation).

  • Assessed random-seed stability (fzgt): We conducted two additional COD10K runs with different random seeds. Across three total runs, FβwF_\beta^w was 0.6917±0.00390.6917 \pm 0.0039, demonstrating stable performance with minimal variance. We also clarified how this stability differs from statistical significance against competing methods.

  • Clarified experimental fairness and hyperparameter selection (Xp3s, fzgt): We detailed our adoption of the DualUCOD protocol (476 × 476 resolution, dataset splits, and evaluation). For competing methods, we standardized the input resolution for fair comparisons and reported DINO and DINOv2-based results separately. Additionally, all hyperparameters were fixed after being tuned on an independent validation subset drawn from COD10K-Train and CAMO-Train.

We thank the reviewers again for their valuable suggestions, which have strengthened the empirical support for DES-UCOD. Detailed results and explanations are provided in our point-by-point responses. We appreciate your time and welcome any further questions.

Best regards,
The Authors of Submission #14625

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.