acceptodds
Withdrawn from ICLR 2027

SegDis: Segment First, Then Discriminate for Unified Salient and Camouflaged Object Detection

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Existing joint learning approaches for salient object detection (SOD) and camouflaged object detection (COD) typically formulate both tasks as binary foreground segmentation, making it difficult to distinguish salient from camouflaged objects when they coexist. We uncover a deeper limitation through bidirectional cross-task transfer: models trained exclusively on SOD generalize surprisingly well to COD, and vice versa. This reveals that strong segmentation performance does not necessarily imply reliable object-type discrimination, suggesting that existing models primarily rely on generic foreground cues rather than task-specific characteristics of saliency and camouflage. Motivated by this observation, we propose Segment First, Then Discriminate (SegDis), a unified framework that explicitly decouples foreground discovery from object-type discrimination. SegDis first identifies task-agnostic foreground regions and then determines whether foreground pixels are salient or camouflaged by conditioning class-specific type queries on the relation between foreground and surrounding background. A subsequent uncertainty–boundary refinement stage resolves ambiguous regions and sharpens object boundaries, yielding a three-class prediction of background, salient objects, and camouflaged objects. On USC12K, SegDis achieves 71.82% FG-mIoU and 80.25% mIoU, outperforming USCNet by 3.36 and 2.22 percentage points, respectively, while reducing the Camouflage-Saliency Confusion Score (CSCS) from 7.49 to 3.27. On Scene-C, where salient and camouflaged objects coexist, SegDis obtains 73.82% IoU-S and 51.60% IoU-C, improving IoU-C by 5.87 percentage points over USCNet. These results demonstrate that decoupling foreground discovery from object-type discrimination improves unified three-class parsing and reduces salient–camouflaged confusion in mixed scenes.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Salient Object Detection, Camouflaged Object Detection, Unified SOD/COD Detection, Foreground Segmentation

Reviews

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

This paper proposes SegDis, a unified framework for salient object detection (SOD) and camouflaged object detection (COD). The method decomposes three-class segmentation into task-agnostic foreground discovery and foreground-conditioned object-type discrimination. It introduces a Progressive Foreground Refinement System (PFRS), relation-conditioned type queries, and an uncertainty-boundary refinement mechanism. Experiments on USC12K demonstrate improvements over USCNet in overall mIoU and foreground mIoU, along with a reduction in saliency-camouflage confusion.

Critical strengths

1.The USC12K results are promising. Compared with USCNet, mIoU improves from 78.03 to 80.25 and CSCS drops from 7.49 to 3.27 (Table 1). The improvement in Scene-C camouflage IoU, from 45.73 to 51.60, is particularly relevant to the problem of detecting both object types in the same image.

2.The separation of foreground segmentation and object-type discrimination provides a reasonable approach to unified SOD and COD, particularly in scenes where both object types coexist.

Critical weaknesses

1.My main concern is how to attribute the largest gain in Table 2. E0 to E1 increases FG-mIoU from 64.874 to 70.867, but changes both the foreground factorization and the query head. Section 4.4 explicitly acknowledges this, so the result supports the combined change. It does not establish how much the factorization itself contributes. E0’s background prediction and losses are also insufficiently specified to assess how comparable the supervision is. A direct three-class head using the same PFRS, comparable head capacity, and matched foreground supervision would help establish the specific benefit of foreground-first prediction. This matters because the factorization is a central contribution of the paper.

2.A smaller concern is the stability of the component gains. Adding relation conditioning in E3 to E5 improves overall FG-mIoU from 71.049 to 71.819 and reduces CSCS from 3.362 to 3.275. Scene-C IoU-S changes from 74.017 to 73.819, and IoU-C from 51.724 to 51.605. These small differences do not establish a benefit or a degradation on Scene-C. The authors disclose that these are single runs and acknowledge the missing seed analysis. The remaining question is whether the aggregate improvement is consistent across runs; the current results provide only limited evidence on this point.

3.The experimental comparison lacks a comparison with some methods from 2025 and 2026, such as SAM2-UNet, VSCodeV2, CamoDiffusion, Samba, which limits the persuasiveness of the assessment.

Questions for authors

1.How are E0’s background prediction and training losses defined, and exactly how do they differ from E1? Is a direct three-class baseline with comparable head capacity and foreground supervision available?

2.Additional experimental evidence: Do you have repeated-run results for E3 and E5? These would help assess whether the gain from relation conditioning is reliable, both overall and on Scene-C.

3.Dividing first and then classifying is an intuitive approach. The technical contributions and innovations shoud be further hightlighed.

4.There is no explanation or discussion of the training strategy, whether it is end-to-end or two-stage training.

Reviewer_Dix2OpenReview ↗
Rating
2
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes SegDis for unified pixel-wise segmentation of background, salient objects, and camouflaged objects. It decomposes prediction into task-agnostic foreground estimation and foreground-conditioned SOD/COD classification. A progressive foreground refinement system estimates foreground support, while pooled foreground–background features condition two learned type queries for pixel-wise classification. A bounded foreground correction and probability composition produce the final three-class output. The complete model is trained jointly without requiring a task identifier at inference.

Critical strengths
  1. Separating foreground localization from type classification is a sensible approach for images containing both salient and camouflaged objects.
  2. The overall gains over USCNet are meaningful, particularly the Scene-C camouflage IoU improvement from 45.73 to 51.60. The failure analysis also helps distinguish localization errors from type confusion.
Critical weaknesses
  1. Table 2 does not isolate the main design choice. E0→E1 changes both the factorization and the prediction head, so it is unclear how much of the gain comes from the proposed decomposition. A comparable direct three-class head using the same PFRS would provide a stronger baseline.
  2. The benefit of relation conditioning is less clear. E3→E5 improves overall mIoU by 0.531 points, but neither Scene-C IoU improves. With only one run and no comparison of conditional type accuracy on Scene-C, its benefit for distinguishing coexisting objects remains uncertain.
Questions for authors
  1. How does a direct three-class head perform with the same PFRS features, comparable head capacity, and auxiliary foreground supervision? Please also clarify E0’s loss and background prediction.
  2. Does relation conditioning improve type accuracy within the ground-truth foreground of Scene-C? Are the E3/E5 differences consistent across repeated runs?
Reviewer_2FtCOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper presents SegDis, a unified framework for salient and camouflaged object detection that separates foreground discovery from pixel-wise object-type discrimination. It uses foreground–background relations to adapt type-specific queries and combines conditional type probabilities with foreground confidence to produce three-class predictions. Experiments on USC12K show improved overall segmentation performance and reduced saliency–camouflage confusion compared with USCNet.

Critical strengths
  1. Clear and interpretable formulation. Separating foreground discovery from type discrimination provides a coherent framework for handling coexisting salient and camouflaged objects without requiring an input task identifier.
  2. Competitive overall performance. SegDis improves USC12K mIoU from 78.03% to 80.25% and reduces CSCS from 7.49 to 3.27 compared with USCNet.
  3. Useful diagnostic analysis. Separate evaluations of foreground segmentation and type discrimination, together with qualitative failure cases, help clarify the model’s behavior and remaining limitations.
Critical weaknesses
  1. The manuscript provides neither anonymous inference code nor trained model weights, making independent verification of the reported results difficult.
  2. No comparisons with competing methods are provided in terms of parameter count, inference speed, or peak GPU memory usage, leaving the trade-off between accuracy and computational cost unclear.
  3. Although SegDis improves the overall metrics over USCNet, performance declines in some key settings, including Scene-B IoU-C and Scene-C IoU-S. Its advantages are therefore not consistent across scenarios and metrics.
  4. The core design closely resembles existing image-conditioned query mechanisms, with the contribution primarily involving the combination and refinement of existing components. Moreover, the ablation studies do not sufficiently isolate the contributions of the proposed core mechanisms.
Questions for authors

Could the authors provide anonymized access to the inference code together with pretrained model checkpoints, to enable reproduction of the reported results?

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

Related papers

Discussion (0)

Sign in to comment.