acceptodds
Under review as a conference paper at ICLR 2027

SEED: Subject-to-Edge Exploration and Distillation for Unsupervised Camouflaged Object Detection

Abstract

Unsupervised Camouflaged Object Detection (UCOD) aims to identify concealed objects without relying on pixel-level annotations. Mainstream methods rely on prior-based pseudo-labels that often misclassify background noise as foreground under high object–background similarity. Recent approaches suppress noise via iterative pseudo-label optimization or uncertaintyaware strategies. However, they face a trade-off between noise filtering and detail preservation: conservative filtering fails to remove noise that disrupts subject localization, whereas aggressive filtering sacrifices edge details for subject purification. To address these issues, we decompose UCOD into a subject localization with distillation phase and a detail exploration phase. Accordingly, we propose a Subject-to-Edge Exploration and Distillation (SEED) method based on the TeacherStudent framework, which comprises three modules: the Subject Teacher Initialization (STI) module, the Subject Consistency Distillation (SCD) module, and the Active Detail Exploration (ADE) module. For subject localization with distillation, STI establishes a purified camouflaged subject via dynamic pseudo-label, then SCD employs confidence-partitioned distillation to maintain subject consistency and mitigate confusion. For detail exploration, ADE leverages temperature rescaling to actively refine details for fine-grained prediction. Extensive experiments demonstrate that our method achieves state-of-the-art performance.

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

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

This paper proposes SEED, a teacher-student framework for unsupervised camouflaged object detection (UCOD), which decouples the task into two stages: “subject localization + distillation” and “detail exploration,” aiming to simultaneously address the trade-off between suppressing noise from false labels and preserving edge and structural details.

Critical strengths

The model designed by the author performs very well and has achieved excellent results on multiple datasets.

Critical weaknesses

The research motivation and technical novelty of this work are relatively limited. The paper attributes the main challenge to the trade-off between “suppressing noise in pseudo-labels” and “preserving fine-grained details.” However, this is not a problem unique to unsupervised camouflaged object detection (UCOD), but rather a common challenge in unsupervised learning involving pseudo-labels, self-training, and teacher-student models. Therefore, the research motivation itself does not reveal a sufficiently novel or insightful problem definition. More importantly, the proposed framework appears to consist primarily of existing, commonly used techniques. The teacher-student paradigm, exponential moving average (EMA)-based pseudo-label optimization, confidence-based region segmentation, consistency distillation, skeletonization, and temperature-based sharpening are all well-established technical components. The main contribution of this paper appears to lie in combining these existing modules into a coarse-to-fine training workflow, rather than introducing some fundamentally new learning mechanism or technical formulation. Although the experimental results indicate that this specific combination is effective, the paper fails to adequately justify why this combination constitutes a substantial methodological advancement over prior work. Furthermore, I have concerns regarding the fairness and completeness of the experimental comparisons. The paper does not explicitly report the input resolution used in its main experimental setup, making it difficult to determine whether the reported performance improvements were achieved under comparable computational conditions. This is particularly critical because segmentation performance is often highly sensitive to image resolution. Furthermore, the comparison with the newer UCOD method is insufficiently comprehensive, making it difficult to establish the method’s position relative to the latest state-of-the-art (SOTA). Overall, this paper presents a well-engineered system that achieves competitive empirical performance; however, its core methodology consists primarily of a combination of existing components, offering limited conceptual novelty. In its current form, I believe its technical contribution is not strong enough and lacks sufficient distinctiveness to meet the level of novelty typically expected at ICLR.

Questions for authors
  1. The central motivation of this paper—the trade-off between pseudo-label noise suppression and detail preservation—is a very common problem in unsupervised learning and self-training. Could the authors further explain what new technical challenges this problem poses in UCOD that have received little attention in previous research?
  2. SEED primarily consists of existing techniques, such as the teacher-student architecture, EMA pseudo-label refinement, confidence segmentation, consistency distillation, soft skeletalization, and temperature sharpening. Could the authors clearly articulate its core technical innovations, rather than merely describing the combination of these existing modules and the reorganization of the training process?
  3. The model size and computational cost of the full SEED (student model) are significantly higher than those of major baseline models (such as UCOD-DPL), and the paper does not explicitly report the input resolutions for all methods in the main experiments. Could the authors conduct a fair comparison under the same architecture, input resolution, and computational budget?
  4. The “Related Work” section of the paper mentions newer UCOD methods such as EReCu and Self-Anchored, but the main experiments do not include direct comparisons with these state-of-the-art methods. In the absence of these strong baseline models, how do the authors support their conclusion that their method is state-of-the-art and offers advantages?
2 comments

Our paper received positive feedback for its clear motivation and strong performance gains. We thank all reviewers for their valuable and encouraging comments, including “an important issue in UCOD,” “The problem statement matches the results,” and “excellent results on multiple datasets.” We address all concerns point by point below.

Q1: Task-Specific Motivation and Challenge

In generic unsupervised segmentation, the trade-off between denoising and detail preservation mainly stems from pseudo-labels’ limited ability to represent semantic and instance structure; in UCOD, it stems from the high foreground–background feature similarity caused by camouflage. This confusion makes denoising more likely to erode genuine object structures or even retain a non-camouflaged object as the target, making the problem more severe in UCOD. Existing methods complete camouflaged object structures through feature-response-based pseudo-label purification and reliable-region selection. Biased feature responses in camouflage scenes leave genuine structures with weak responses without effective supervision, while non-camouflaged objects with strong responses are incorrectly retained. To address the lack of supervision for genuine details, we construct reliable subject constraints and detail supervision separately. Under reliable subject constraints, we use structural continuity and local relative responses to explore potential details, reducing the dependence of detail recovery on the completeness of teacher masks.

Q2: Design Rationale and Core Contribution of SEED

Existing UCOD methods rely on high-level semantic features from self-supervised visual models but lack fine-grained detail. To recover details, we directly introduced multi-level features into these methods but observed severe performance degradation. These results indicate that noisy pseudo-labels limit the effective use of fine-grained information across feature layers. This observation points to a mismatch between fine-grained representational capacity and the reliability of pseudo-label supervision.

SEED addresses this mismatch by decoupling subject localization from detail exploration. The teacher uses high-level semantic features to establish reliable subject-level constraints. Guided by these constraints, the student actively explores potential details using spatial information across feature layers. This organization is designed to reduce the interference of noisy supervision with fine-grained feature learning and to make detail recovery less dependent on complete teacher masks.

SEED’s innovation lies not in a specific architecture, but in reorganizing learning objectives and training around the pseudo-label reliability requirements of different features. SEED is therefore not a simple assembly of existing components into a coarse-to-fine training pipeline.

Q3: Controlled Comparison under Matched Model Capacity

We thank the reviewer for raising the question of experimental fairness. We added a controlled comparison in which SEED (DBA+LT) and UCOD-DPL use the same model architecture, input features, Look-Twice (LT) strategy, and 518×518518×518 input resolution. The first SEED round consists of 20 epochs of STI and 5 epochs of SCD+ADE, matching UCOD-DPL's total of 25 training epochs. We also evaluated a single 1×11×1 convolutional decoder under the same experimental settings to examine the role of decoder capacity.

DatasetRoundConfigurationSm ↑Fw ↑meanFm ↑meanEm ↑MAE ↓
COD10K—UCOD-DPL.834.763.779.916.0310
Round 1SEED (1×1).829.750.775.906.0261
Round 1SEED (1×1 + LT).843.784.811.912.0248
Round 1SEED (DBA).831.755.779.912.0266
Round 1SEED (DBA+LT).847.788.814.918.0250
Round 2SEED (DBA).834.757.774.915.0254
Round 2SEED (DBA+LT).855.797.816.925.0235

With the matched DBA+LT configuration, first-round SEED improves all five reported metrics over UCOD-DPL. The lower-capacity 1×11×1+LT configuration also improves Sm, Fw, meanFm, and MAE, although its meanEm is lower than that of UCOD-DPL. These results support the effectiveness of the proposed training strategy under matched or lower decoder capacity, without requiring a second round.

More generally, decoder complexity does not guarantee better performance in an unsupervised setting. Greater expressiveness can also increase the risk of fitting pseudo-label noise. The heterogeneous decoder design is motivated by the distinct requirements of subject localization and detail exploration.

Q4: Comparisons with Recent UCOD Methods

SEED outperforms these methods. Please see our response to Reviewer kZAH Q3 due to the character limit.

The authors adopt a 518×518 input resolution that is consistent only with UCOD-DPL but not with the other compared methods, which raises a serious concern about experimental fairness, as it remains unclear whether the reported gains stem from the higher resolution rather than from the effectiveness of the proposed method; I would suggest repeating the comparisons under a unified resolution or providing an ablation study to isolate the effect of resolution. Furthermore, the novelty of this work still appears limited, since the proposed approach shares substantial technical similarities with UCOD-DPL and reads more as a technical extension of it without introducing a genuinely new theory or algorithm. In light of these concerns, I maintain my original score.

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

SEED studies unsupervised camouflaged object detection. The authors argue that filtering noise in a single pseudo-label leads to a trade-off between removing background clutter and preserving edges, and they split the task into subject localization and detail exploration. The teacher–student framework has three modules. STI trains the teacher with a dynamic mix of a fixed prior and an EMA prediction to obtain a clean subject. SCD distills teacher–student consistency on high-confidence regions and suppresses hesitant predictions on confused regions. ADE applies temperature rescaling inside skeleton-connected regions and uses the result to supervise the student. On CHAMELEON, CAMO, COD10K, and NC4K, the authors report an average Sm gain of about 2.5% over UCOD-DPL, and a 23.02% reduction in MAE.

Critical strengths
  1. The problem statement matches the results. In Table 2, round 2, adding only SCD or only ADE moves COD10K Sm from .840 to .843 and .841, while using both reaches .859. Either module alone does almost nothing, and the two together help. This interaction supports the claim that enforcing subject consistency alone smooths boundaries, while sharpening alone amplifies noise.

  2. Boundary quality is measured separately. BIoU in Table 4 is higher than UCOD-DPL on all four datasets, for example COD10K .535 to .591 and NC4K .503 to .587. The gain shows up on boundaries, not only in region error.

  3. The appendix checks the asymmetric decoders, feature levels, training rounds, and the sensitivity of τce, K, and temperature T. Table 6 reports that some metrics drop after the third round. Table 7 evaluates pseudo-label quality against training-set ground truth that is never used for training, and the quality improves. On COD10K and NC4K, all five metrics move in the same direction relative to UCOD-DPL.

Critical weaknesses
  1. Relative to UCOD-DPL, the new pieces are mainly alignment to the teacher on high-confidence regions and sharpening predictions on the skeleton. The dynamic pseudo-label, EMA, FOUND prior, and DBA decoder follow that prior work. Table 2 shows that the two new modules together give about +0.019 Sm, but this is measured with the larger decoder. The main-table jump from UCOD-DPL’s .834 to .859 is +0.025, and part of it may come from the larger decoder and the second training round. The best reported student uses 14.74M trainable parameters and 425.59G MACs, against 0.10M and 117.5G for UCOD-DPL. With the same small decoder, Table 8 gives .842 Sm in round 1, only +0.008 over UCOD-DPL. The same-architecture teacher reaches .844 in round 2. A round-2 student at matched capacity is not reported, so the +0.025 cannot yet be attributed entirely to the two modules.

  2. Equations (11)–(12) rescale only pixels on connected components of the skeleton and leave all other pixels unchanged. The text describes searching outward from the object and recovering edges, but the equations do not say how the region expands from the skeleton to the boundary. If the implementation dilates the skeleton or uses a neighborhood, that should be stated in the method.

  3. The claimed 23.02% MAE reduction in the abstract is about 18% when computed from Table 1 as (old − new) / old, averaged over the four datasets. The 23% figure is closer to dividing by the new error.

  4. Self-Anchored is discussed in related work and claims to outperform prior unsupervised methods, but it is absent from the main table. Without that comparison, the claim of outperforming existing UCOD methods is incomplete.

Questions for authors
  1. Report the round-2 student, using the same DBA decoder and the same last-layer features as UCOD-DPL, on all four datasets.

  2. Add Self-Anchored to the main comparison table.

  3. Report the mean and standard deviation over at least three random seeds on COD10K.

2 comments

Our paper received positive feedback for its clear motivation and strong performance gains. We thank all reviewers for their valuable and encouraging comments, including “an important issue in UCOD,” “The problem statement matches the results,” and “excellent results on multiple datasets.” We address all concerns point by point below.

Q1: The Effect of Model Capacity on Performance

SEED achieves gains with simple decoders, its gains do not rely solely on a larger decoder or a second training round. All methods use a 518×518 input resolution. Each SEED round consists of 20 epochs of STI followed by 5 epochs of SCD+ADE (25 epochs total). The first SEED round therefore matches UCOD-DPL’s 25 training epochs.

We evaluated SEED with either a DBA decoder or a single 1×1 convolutional decoder, each with and without Look-Twice (LT) test-time augmentation. The reported UCOD-DPL results use LT, so SEED configurations with LT match this test-time setting. Due to the character limit, the table shows only second-round DBA results (with and without LT), alongside first-round 1×1+LT results.

datasetMethodsSm ↑Fw ↑MAE ↓
CHAMELEONUCOD-DPL.864.825.0310
SEED(1×1+LT, R1).875.844.0273
SEED(DBA, R2).865.810.0275
SEED(DBA+LT, R2).880.850.0265
CAMOUCOD-DPL.793.747.0770
SEED(1×1+LT, R1).796.749.0742
SEED(DBA, R2).802.755.0686
SEED(DBA+LT, R2).802.760.0692
COD10KUCOD-DPL.834.763.0310
SEED(1×1+LT, R1).847.788.0248
SEED(DBA, R2).834.757.0254
SEED(DBA+LT, R2).855.797.0235
NC4KUCOD-DPL.850.818.0430
SEED(1×1+LT, R1).854.826.0379
SEED(DBA, R2).860.826.0342
SEED(DBA+LT, R2).868.842.0333

In unsupervised settings, increasing decoder complexity does not necessarily improve performance: a more expressive model can also fit pseudo-label noise. SEED's heterogeneous decoders are designed around the distinct supervision requirements of subject localization and detail exploration, balancing noise filtering with detail recovery.

As for dinov1, STI localizes subjects less accurately than DINOv2. Subsequent consistency constraints and local sharpening cannot fully compensate for missing discriminative information, limiting the gains over UCOD-DPL.

Q2: Comparisons with Recent UCOD Methods

SEED outperforms these methods. Please see our response to Reviewer kZAH Q3 due to the character limit.

Q3: Results over Three Random Seeds

We repeated the experiments using three random seeds and report the mean ± standard deviation on all four datasets below. The observed variation across runs supports the stability of SEED's reported performance.

DatasetSm ↑Fw ↑meanFm ↑meanEm ↑MAE ↓
CHAMELEON.8745 ± .0036.8305 ± .0029.8492 ± .0018.9274 ± .0067.02629 ± .00045
CAMO.8089 ± .0054.7596 ± .0032.7989 ± .0041.8574 ± .0035.07217 ± .00153
COD10K.8557 ± .0045.7968 ± .0014.8248 ± .0031.9288 ± .0056.02336 ± .00043
NC4K.8717 ± .0041.8467 ± .0016.8707 ± .0013.9327 ± .0037.03372 ± .00072

W1: Novelty of SEED’s STI Beyond UCOD-DPL

STI filters noise more aggressively than UCOD-DPL and defers detail recovery to later stages. This separation reduces the risk of learning background noise as foreground and reinforcing pseudo-label errors. UCOD-DPL balances noise filtering and detail preservation within the same process, limiting its ability to match STI’s subject purification even with relatively clean pseudo-labels.

To further examine this distinction, we used the Round 1 SEED student predictions as the pseudo-supervision source for continued training with UCOD-DPL and with STI.

MethodSm ↑weighted F ↑max F ↑max E ↑MAE ↓
UCOD-DPL.832.765.796.912.0308
STI Teacher.844.780.799.921.0270

Despite reusing UCOD-DPL components, the STI teacher achieves better subject localization under the same pseudo-supervision, indicating distinct optimization behavior.

W2: Clarification of the Skeleton-Guided Exploration Regions

The skeleton guides the selection of local exploration patches. Temperature rescaling is applied to all pixels within the selected patches, rather than only to skeleton pixels. The exploration regions therefore cover the areas around the skeleton through patch coverage; they are not obtained by explicitly dilating the skeleton itself.

W3: Correction to the Relative MAE Reduction

We used SEED’s MAE instead of UCOD-DPL’s as the denominator. The corrected average relative MAE reduction is 18.25%. We will update the manuscript.

Thank you for the additional experiments and clarifications. The DBA+LT result is particularly helpful: obtaining .855 Sm on COD10K with the smaller decoder substantially reduces my concern that SEED’s gain depends primarily on the larger student decoder. I also appreciate the additional ADE ablation and the clarification that the skeleton guides patch selection rather than serving as a boundary map. I think the paper could be made more informative by examining what the skeleton contributes to this patch selection. The new comparison shows that global sharpening alone and centering alone do not reproduce the full ADE result, but the configurations differ in more than the use of skeleton guidance. A useful controlled comparison would keep the patch size, number or coverage budget, centering and temperature fixed, while selecting patches using, for example, random locations or prediction uncertainty instead of the skeleton. This would help distinguish whether the gain comes from skeleton-guided localization specifically or from applying the refinement operation locally. Finally, given the error identified in the reported average MAE reduction, I encourage the authors to conduct a thorough consistency check of all numerical results in the main paper and supplementary material. In particular, please verify the values against the underlying experimental outputs, recalculate derived statistics and relative improvements, and check that results and experimental settings are reported consistently across tables, figures, and the text.

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

This paper proposes SEED (Subject-to-Edge Exploration and Distillation), a teacher–student framework for unsupervised camouflaged object detection (UCOD). The main motivation is to address the trade-off between suppressing noisy pseudo-labels and preserving fine-grained object boundaries.

SEED divides the learning process into subject localization and detail exploration. It introduces three components: Subject Teacher Initialization (STI), which employs dynamic pseudo-labels to initialize a teacher model; Subject Consistency Distillation (SCD), which partitions predictions into high-confidence and ambiguous regions for selective distillation and confusion suppression; and Active Detail Exploration (ADE), which employs skeleton-guided local temperature rescaling to refine pseudo-labels and recover fine structural details.

Experiments are conducted on CHAMELEON, CAMO-Test, COD10K-Test, and NC4K. The proposed method improves S-measure from 0.834 to 0.859 on COD10K-Test and from 0.850 to 0.876 on NC4K compared with UCOD-DPL. The authors also provide module ablations, backbone comparisons, boundary-quality evaluations, and several additional analyses.

Critical strengths
  1. A well-designed coarse-to-fine framework for pseudo-label refinement.

The paper addresses an important issue in UCOD: filtering noisy pseudo-labels can improve training, but it may also remove useful object details. Separating object localization from detail refinement is a sensible way to handle this trade-off. The ablation results in Table 2 also support this design. On COD10K-Test, the second-round model achieves an S-measure of 0.843 with STI and SCD, and 0.841 with STI and ADE. When all three modules are combined, the score increases to 0.859. These results suggest that the modules complement each other and that the full pipeline benefits from their combination.

  1. Strong results on UCOD benchmarks and boundary quality.

The proposed method shows consistent improvements over UCOD-DPL on both COD10K-Test and NC4K. Beyond the gains in S-measure and MAE, the improvement in boundary IoU is especially worth noting, as boundary preservation is one of the main goals of the paper. As reported in Table 4, boundary IoU improves from 0.535 to 0.591 on COD10K and from 0.503 to 0.587 on NC4K. These are solid improvements, and the qualitative examples and ablation studies provide further support for the method's effectiveness.

Critical weaknesses
  1. Limited control for model capacity and computational cost.

It is unclear how much of the performance gain comes from the proposed training strategy rather than the larger student decoder. As shown in Table 5, SEED uses 14.74M trainable parameters and 425.59G MACs, compared with only 0.10M and 117.5G for UCOD-DPL. Although the decoder ablations in Table 8 are helpful, a comparison using the same decoder and training setup would be more convincing.

The improvement with DINOv1 is also quite small. Table 3 shows that S-measure increases only from 0.727 to 0.729, while E-measure slightly decreases. Since all experiments use the same random seed, it is difficult to tell whether these small differences are reliable. More controlled experiments would help clarify how much the proposed training strategy contributes to the overall improvement.

  1. Insufficient validation of skeleton-guided detail exploration.

The paper does not provide enough evidence that skeleton guidance is actually responsible for ADE's improvements. Eq. (12) combines local logit-centering with temperature scaling, but the effects of these two operations are not evaluated separately. The connection between the skeleton-based exploration regions and object boundaries is also not clearly explained.

More importantly, the paper lacks comparisons with simpler alternatives, such as global temperature sharpening, local sharpening without skeleton guidance, and standard boundary refinement. Without these baselines, it is hard to tell whether the improvement comes from skeleton guidance or simply from sharpening the predictions. This weakens the claim that ADE introduces a meaningful new mechanism.

Questions for authors
  1. How much of SEED's improvement comes from the proposed training strategy rather than model capacity?

Can the authors compare SEED with UCOD-DPL using the same BiRefBlock decoder and hierarchical DINOv2 features, under a similar training budget? Results from multiple random seeds would also help confirm the reliability of the improvements. In addition, why is the improvement with DINOv1 so small?

  1. What does skeleton guidance actually contribute to ADE?

According to Eq. (12), setting T=1 does not recover the original probability because the local logit-centering term is still present. Does the implementation skip this transformation when T=1?

Also, how are skeleton-connected components used to define the exploration regions? Can the authors compare ADE with simpler methods, such as global or local temperature sharpening without skeleton guidance? These comparisons are needed to show whether skeleton guidance itself is responsible for the improvement.

  1. How is the reported 23.02% average MAE reduction calculated?

The paper reports an average MAE reduction of 23.02% over UCOD-DPL. However, based on the results in Table 1 for CHAMELEON, CAMO-Test, COD10K-Test, and NC4K, the average relative reduction appears to be around 18.25%. Can the authors explain how the reported 23.02% was calculated?

  1. Why are recent UCOD methods missing from the main comparison?

The paper discusses recent methods such as EReCu and Self-Anchored, but neither is included in the main quantitative comparison. Why were these methods excluded? Can the authors provide results under the same evaluation protocol to better support the claimed state-of-the-art performance?

1 comment

Our paper received positive feedback for its clear motivation and strong performance gains. We thank all reviewers for their valuable and encouraging comments, including “an important issue in UCOD,” “The problem statement matches the results,” and “excellent results on multiple datasets.” We address all concerns point by point below.

Q1: Controlled Comparison under Matched Model Capacity

The controlled comparison supports the contribution of SEED's training paradigm beyond decoder capacity. In an unsupervised setting, a more expressive decoder can also amplify background textures and incorrect boundaries in noisy pseudo-labels. We replaced the UCOD-DPL decoder with the same BiRefBlock decoder and hierarchical DINOv2 features used by the SEED student. This replacement did not improve UCOD-DPL, as shown below.

DatasetMethodSm ↑Fw ↑meanEm ↑MAE ↓
CHAMELEONUCOD-DPL + BiRefBlock.464.126.341.131
SEED + BiRefBlock.877.831.931.026
CAMOUCOD-DPL + BiRefBlock.385.150.343.296
SEED + BiRefBlock.812.762.858.071
COD10KUCOD-DPL + BiRefBlock.483.166.411.130
SEED + BiRefBlock.859.798.934.023
NC4KUCOD-DPL + BiRefBlock.457.184.398.187
SEED + BiRefBlock.876.848.935.033

SEED assigns complementary roles to its asymmetric teacher and student. The lightweight DBA teacher uses high-level semantic features to obtain reliable subject localization, while the BiRefBlock student uses hierarchical features to recover details under reliable subject-level constraints. The results in Appendix B, Tables 8–9, further support the importance of reliable supervision when using a more expressive decoder for detail recovery. The asymmetric design therefore supports the division of responsibilities between subject localization and detail exploration.

Q2: Ablation of Logit Centering and Temperature Scaling

Eq. (12) combines local logit centering with temperature scaling. Our implementation retains centering when T=1, so we label this configuration “w/ centering, T=1” to distinguish it from removing the entire ADE stage (w/o centering, T=1), as reported in Table 2. We additionally evaluate temperature scaling alone over the entire prediction map as a global sharpening baseline (w/o centering, T=.3). The Round 2 results are shown below.

Second-round configurationSm ↑Fw ↑meanFm ↑meanEm ↑MAE ↓
w/o centering, T=1.843.771.797.914.026
w/o centering, T=.3T=.3.834.750.785.905.028
w/ centering, T=1T=1.843.782.819.907.026
w/ centering, T=.3T=.3.859.798.823.934.023

We observe that some genuine object details have weak foreground responses but remain spatially continuous. Without direct supervision for these details, selecting responses by magnitude alone fail to recover them. ADE uses soft skeletonization to extract linear response structures from the continuous prediction map as candidate detail cues. These structures guide patch selection, and rescaling is applied to all pixels in the selected patches. The skeleton thus provides a cue for selecting exploration regions rather than an explicit boundary map. Neither global temperature scaling alone nor local logit centering alone reproduces the gains of full ADE.

Q3: Correction to the Relative MAE Reduction

We used SEED’s MAE instead of UCOD-DPL’s as the denominator. The corrected average relative MAE reduction is 18.25%. We will update the manuscript.

Q4: Comparisons with Recent UCOD Methods

We agree that the comparison should include these recent methods. EReCu and SAPNet (Self-Anchored) were initially omitted from the main quantitative comparison because their public implementations were incomplete. We now include them together with DualUCOD in the COD10K comparison below, covering DINOv2-based settings.

For DualUCOD, we report the result from its paper because the incomplete core implementation prevented reproduction. For SAPNet, we implemented the pipeline described in its paper and adapted it to DINOv2. The table includes the resulting SAPNet scores, including its poor performance under this adaptation.

MethodSm ↑E ↑F weighted ↑MAE ↓
SAPNet* (TIP 2026)0.3530.4870.0320.267
DualUCOD (ICML 2026)0.7610.8560.6350.042
EReCu* (CVPR 2026)0.7940.8900.7080.041
UCOD-DPL* (CVPR 2025)0.8340.9160.7630.031
SEED (R1-Teacher)0.7840.8690.6750.036
SEED (R1-Student)0.8500.9270.7790.026
SEED (R2-Teacher)0.8440.9210.7800.027
SEED (R2-Student)0.8590.9340.7980.023

* denotes results obtained from our own experiments

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.