Pith. sign in

REVIEW 3 major objections 5 minor 66 references

Segment Concealed Objects with Incomplete Supervision

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A single SAM-based mean-teacher framework, SEE, is the first unified method for incompletely supervised concealed object segmentation, beating specialized baselines across six tasks.

desk verdict Solid extension of WS-SAM into a unified weakly/semi-supervised framework with strong empirical breadth; SOTA claim needs a code release and error bars before I'd trust the margins. read the letter →

arxiv 2506.08955 v2 pith:WJCDJC7F submitted 2025-06-10 cs.CV cs.AIcs.LG

Chunming He , Kai Li , Yachao Zhang , Ziyun Yang , Youwei Pang , Longxiang Tang , Chengyu Fang , Yulun Zhang
show 3 more authors
Linghe Kong Xiu Li Sina Farsiu
This is my paper
classification cs.CVcs.AIcs.LG
keywords concealedobjectsegmentationincompletesupervisionweaklysupervisedsemi-supervisedpseudo-labelingSegmentAnythingModelmean-teacherfeaturegrouping
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper tries to establish that one framework, SEE, can train concealed-object segmenters from incomplete labels — a few points, scribbles, or partially labeled images — as well as or better than methods built separately for each setting. It claims SEE is the first unified method for incompletely supervised concealed object segmentation (ISCOS), reaching state-of-the-art results on weakly and semi-supervised camouflaged object detection, polyp image segmentation, transparent object detection, and salient object detection. SEE wraps a mean-teacher loop around the Segment Anything Model (SAM): the teacher's coarse mask is turned into point, box, and mask prompts, SAM converts those into dense pseudo-labels, and entropy-based storage and selection decide which labels and which pixels supervise the student. A second module, HGFG, groups features at two granularities to make segmentation more coherent, and the paper reports that both SEE and HGFG can be plugged into existing segmenters to improve them.

What carries the argument

The central object is SEE, an iterative mean-teacher framework in which the teacher, updated by exponential moving average, produces coarse fused masks from K augmented views; multi-density prompt extraction (MDPE) then derives foreground and background point prompts via a nine-block sampling strategy, an adaptively expanded bounding box, and a filtered mask prompt, which are fed to SAM to generate dense masks that are fused into pseudo-labels. An optimal label pool keeps the top-B best pseudo-labels according to three entropy-based uncertainty scores (absolute, relative, and differential), and supervision applies an indicator-based image selection plus an entropy-derived pixel weighting so the student trains mostly on reliable regions. The supporting module is HGFG, which groups image features toward N learnable cluster prototypes using iterative GRU updates, then aggregates two granularities (N1=2 and N2=4) with a learned gate in an RK2-style residual combination to promote both within-object and across-object feature coherence.

What would settle it

Train SEE exactly as described but replace SAM's fused pseudo-label with the teacher's coarse fused mask, or with random masks of matched confidence; if segmentation performance does not drop, the SAM stage is not load-bearing. Separately, on a held-out subset with ground truth, compare the F-beta of pseudo-labels kept by the optimal pool against those discarded each epoch: if discarded labels are consistently more accurate, the entropy-based selection rule is selecting for the wrong signal.

Watch

Extended reading notes

Core claim

The central claim is that incomplete supervision is not a separate problem for each concealed-object task: a single SAM-based mean-teacher pipeline can generate pseudo-labels reliable enough to train a student segmenter to top performance. The pipeline's load-bearing elements are multi-augmentation fusion of teacher masks, multi-density prompt extraction (points, an adaptively expanded box, and a filtered mask) to guide SAM, a small pool that stores the best pseudo-labels seen so far, and entropy-based image-level selection plus pixel-level weighting that keeps the student focused on trustworthy regions. The paper also claims that hybrid-granularity feature grouping, HGFG, which clusters features toward learnable prototypes at two granularities and aggregates them with a weighted Runge-Kutta-style gate, fixes incomplete and multi-object segmentation by enforcing feature coherence. Together these components outperform the previous state of the art across six ISCOS tasks, and the authors argue SEE works orthogonally to existing methods, so adding it to them improves their results.

Load-bearing premise

The whole pipeline assumes that SAM, prompted with a coarse teacher mask's points, expanded box, and filtered mask, produces pseudo-labels more accurate than the teacher's own mask, and that pixel entropy of the fused pseudo-label is a reliable proxy for label error.

Editorial extensions

If this is right

  • A single training recipe covers both weak and semi supervision, so users of concealed-object segmentation no longer need separate algorithms for point, scribble, and partially labeled data.
  • SAM's prompt mechanism can be driven by a teacher network's coarse outputs rather than by human annotations, extending a vision foundation model to concealed scenes where it otherwise struggles.
  • Entropy-based pseudo-label storage and supervision should counteract confirmation bias, since the student is trained mostly on regions where the pseudo-label is confident and informative.
  • Because HGFG also improves fully supervised segmenters, feature grouping itself is a generalizable component rather than a fix specific to incomplete supervision.
  • Existing weakly and semi-supervised COD, polyp, glass, and salient object segmenters can be upgraded by inserting SEE or HGFG, per the paper's plug-and-play experiments.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A testable extension beyond the paper's fixed thresholds is to make the entropy selection rules adaptive per dataset or per object size; the paper itself flags small and multi-instance concealed objects as a current limitation.
  • The framework's success likely depends on how well SAM's promptable masks complement the teacher's coarse masks; a direct probe would be to replace SAM's output with a similarly sized random or teacher-only mask and measure the drop, isolating what SAM contributes.
  • The optimal label pool's quality metric (absolute, relative, and differential uncertainty) is validated mainly on COD10K; checking whether pooled pseudo-labels actually have higher F-beta than discarded ones on polyp and glass datasets would test whether the storage rule generalizes.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper proposes SEE, a unified mean-teacher framework for incompletely supervised concealed object segmentation (ISCOS), covering both weakly supervised and semi-supervised settings across camouflaged object detection, polyp segmentation, transparent object detection, and salient object detection. The method uses SAM to generate pseudo-labels from multi-augmentation fused teacher masks, extracts multi-density prompts (points, box, mask), stores the best pseudo-labels in an optimal pool, and applies entropy-based image-level selection and pixel-level weighting for supervision. A hybrid-granularity feature grouping (HGFG) module is introduced to improve feature coherence and handle multi-object cases. Extensive experiments on six tasks report state-of-the-art results over prior methods, with additional plug-and-play integration experiments on existing segmenters.

Significance. If the reported results are reliable, this is a substantial contribution: a single SAM-based pseudo-labeling framework that unifies weak and semi-supervised training for several concealed-object segmentation tasks, with a plug-and-play feature grouping module that boosts the performance of existing models. The paper is strong in breadth: it compares with many specialized methods on multiple benchmarks, includes ablation studies for nearly every design choice, and provides visualizations of feature grouping behavior. The main weakness is that the state-of-the-art claim is not fully evidenced because key hyperparameters are selected on COD10K, which also appears in the final comparison tables, and no repeated-run statistics are reported for the small performance margins over the closest competitor WS-SAM. These issues are fixable and do not invalidate the overall approach, but they need to be addressed before the SOTA claim can be accepted.

major comments (3)
  1. [IV-D, Tables V–XIII vs. Tables I, III] The ablation studies in Section IV-D are stated to be conducted 'on COD10K', and these ablations are used to set K, B, tau_a, tau_r, C, N1, N2, T, and the HGFG placement. The same COD10K test split appears in the main comparison tables (Table I and Table III). If the ablation numbers were computed on the COD10K test split, then the COD10K rows in Tables I and III are not an independent test of the method, and the reported hyperparameters are partially selected on test data. The paper should specify whether a held-out validation split (or the training split) was used for the ablations; if the test split was used, the COD10K comparisons must be re-run with hyperparameters selected without access to the test split.
  2. [Tables I, III, IV] No error bars, standard deviations, or repeated-run statistics are reported anywhere, although several of the claimed gains over WS-SAM are very small. For example, in Table I (point supervision) SEE vs. WS-SAM on CHAMELEON differs by 0.001 in S_alpha (0.806 vs. 0.805) and 0.001 in M (0.055 vs. 0.056); similarly small gaps appear on CAMO. In the absence of seed-level variance, it is not possible to determine whether these differences are systematic. Please report mean and standard deviation over at least three random seeds for the main comparisons, or at least for the SEE-vs-WS-SAM pairs, and state whether the differences are statistically significant.
  3. [Abstract and Section V] The abstract claims that the approach achieves 'complete segmentation for both single-object and multiple-object images' and 'robust network training', but Section V explicitly states that the method has challenges generalizing to 'concealed objects that are small or involve multiple instances'. This is a direct internal inconsistency. The abstract and conclusions should be tempered to match the stated limitations, e.g., by claiming improved completeness in typical cases while acknowledging the small-object and multi-instance limitations.
minor comments (5)
  1. [Section I (contributions) and Section IV-A] The manuscript claims 'the first unified method for ISCOS' and 'the first attempt to leverage the vision foundation model to address ISCOS tasks'. Given that the authors' own WS-SAM [11] is a prior conference version and CamoTeacher [12] is a semi-supervised COD method, the 'unified' claim is defensible only if it covers both weak and semi-supervised settings across tasks; please add a sentence explicitly positioning SEE relative to these prior works to avoid overclaiming novelty.
  2. [Header and Section IV-A] The paper header says 'Official Code TPAMI Version' but no code repository URL is provided anywhere. Please include the URL in the abstract or implementation details section for reproducibility.
  3. [Eq. (4) and Eq. (7)] The entropy E(·) is computed with base 2, so its values lie in [0,1] for probability inputs in [0,1]. The threshold 0.9 and the weighting (1−E) are then well-defined, but the paper should explicitly state that the inputs to E are the soft pseudo-label probabilities (not hard 0/1 values) and that 0.9 is a relative threshold on the maximum entropy.
  4. [Table VIII] The column heading 'SA & W A' contains an extra space; it should read 'SA & WA'. The same typo appears in the caption of Table XV and Table XVI ('Restuls' for 'Results').
  5. [Algorithm 1] In Algorithm 1, line 9 stores the top-B best labels at each epoch, but the comparison rule (at least two of three metrics) is described only in the text. Adding a one-line comment in the algorithm or pointing to Eq. (5) more explicitly would improve readability.

Circularity Check

1 steps flagged · score 6.0 of 10

COD10K test-set tuning makes that benchmark partially fitted; otherwise the derivation is self-contained.

  1. fitted input called prediction [Section IV-D (Ablation Study), Section IV-A (Datasets); Tables I and III]
    "We conduct ablation studies on COD10K of the COD task in the weak supervision with scribble. ... our training set comprised 1,000 images from CAMO and 3,040 images from COD10K, with the remaining images serving as the test set. ... Table XI indicates that the best settings for the absolute uncertainty threshold and relative uncertainty threshold are τ_a = 0.1 and τ_r = 0.5, respectively."

    The hyperparameters K, B, τ_a, τ_r, the point count C, and HGFG placement are chosen by ablations run on the COD10K split, and the same split is then reported as a held-out test result in Tables I and III. Consequently, the COD10K entries in the main tables are not independent predictions but in-sample values obtained after fitting these hyperparameters to that benchmark. The SOTA claim retains independent support on the non-tuned benchmarks (CHAMELEON, CAMO, NC4K, polyp, TOD, SOD), so the circularity is partial rather than total.

full rationale

Apart from the COD10K benchmark-contamination issue, the derivation chain is self-contained. The pseudo-label generation, storage, and supervision operators (Eqs. 1-8) are training heuristics whose reported test numbers come from external benchmarks; no equation reduces by construction to its own target metric. The reliance on the authors' earlier WS-SAM is an explicit baseline and extension, not a load-bearing self-citation chain, and no uniqueness theorem is imported from the authors' prior work. HGFG is a feature-grouping module with external inspiration (slot attention) and is validated by ablation. Therefore, the only concrete reduction of a 'prediction' to fitted inputs is the COD10K test-set selection of hyperparameters.

Assumptions & free parameters 9 free parameters · 4 assumptions · 0 invented entities

The claim rests on empirical engineering choices rather than on a derivation. The free parameters are the augmentation count, pool size, selection thresholds, point count, prototype counts, and EMA momentum. The axioms are assumptions about SAM's transfer, entropy as a proxy for label quality, EMA stability, and the benefit of feature clustering. No new physical entities are introduced.

free parameters (9)
  • K: number of augmented views = 12
    Sets the ensemble size for multi-augmentation fusion in Eq. (2); chosen via Table XI on COD10K.
  • B: number of stored best pseudo-labels = 3
    Optimal pool size in Eq. (5); Table XI shows diminishing returns beyond 3.
  • tau_a: absolute uncertainty threshold = 0.1
    Image-level selection threshold in Eq. (6); selected via Table XI.
  • tau_r: relative uncertainty threshold = 0.5
    Second image-level selection threshold in Eq. (6); selected via Table XI.
  • High-uncertainty entropy threshold = 0.9
    Pixels with entropy above 0.9 are treated as high-uncertainty in the pseudo-label storage and supervision criteria (Section III-A2).
  • C: number of key points = 9
    Nine-block strategy samples 9 points; set after Fig. 9 analysis.
  • N1, N2: prototype counts in HGFG = 2, 4
    Hybrid granularity grouping in Eq. (13); selected via Table XIII.
  • T: GRU iterations in feature grouping = 3
    Iterations of prototype updates in Eq. (12); selected via Table XIII.
  • EMA momentum eta = 0.996
    Teacher update in Eq. (18); standard choice, not ablated.
assumptions (4)
  • domain assumption SAM (ViT-H) can produce useful dense masks for concealed objects when given teacher-derived point, box, and mask prompts.
    Invoked at Eq. (3); if SAM ignores or is misled by coarse prompts, the pseudo-labels inherit teacher errors. The authors themselves note SAM struggles in concealed scenes (Section III-A), so prompting is assumed to fix this.
  • ad hoc to paper Pixel entropy of a fused pseudo-label is a reliable proxy for label correctness and usefulness.
    Used in Eq. (4), Eq. (6), and Eq. (8) for storage, selection, and weighting; no validation against ground-truth error rates is provided.
  • domain assumption EMA distillation keeps the teacher model accurate enough for its coarse masks to improve over training.
    Eq. (18) assumes the standard mean-teacher dynamics; early-stage pseudo-labels are known to be noisy (Section I), so the selection pipeline must compensate.
  • domain assumption Grouping features into a small number of learned prototypes improves feature coherence and reduces incomplete segmentation.
    Section III-B and Eq. (9)-(13) posit this clustering benefit; ablations support it, but the mechanism is not proven for concealed objects.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Segment Concealed Objects with Incomplete Supervision." pith.science (2026). https://pith.science/paper/WJCDJC7F

@misc{pith2026250608955,
  author       = {Pith},
  title        = {Pith review of: Segment Concealed Objects with Incomplete Supervision},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WJCDJC7F}},
  note         = {Machine review of arXiv:2506.08955}
}
read the original abstract

Incompletely-Supervised Concealed Object Segmentation (ISCOS) involves segmenting objects that seamlessly blend into their surrounding environments, utilizing incompletely annotated data, such as weak and semi-annotations, for model training. This task remains highly challenging due to (1) the limited supervision provided by the incompletely annotated training data, and (2) the difficulty of distinguishing concealed objects from the background, which arises from the intrinsic similarities in concealed scenarios. In this paper, we introduce the first unified method for ISCOS to address these challenges. To tackle the issue of incomplete supervision, we propose a unified mean-teacher framework, SEE, that leverages the vision foundation model, ``\emph{Segment Anything Model (SAM)}'', to generate pseudo-labels using coarse masks produced by the teacher model as prompts. To mitigate the effect of low-quality segmentation masks, we introduce a series of strategies for pseudo-label generation, storage, and supervision. These strategies aim to produce informative pseudo-labels, store the best pseudo-labels generated, and select the most reliable components to guide the student model, thereby ensuring robust network training. Additionally, to tackle the issue of intrinsic similarity, we design a hybrid-granularity feature grouping module that groups features at different granularities and aggregates these results. By clustering similar features, this module promotes segmentation coherence, facilitating more complete segmentation for both single-object and multiple-object images. We validate the effectiveness of our approach across multiple ISCOS tasks, and experimental results demonstrate that our method achieves state-of-the-art performance. Furthermore, SEE can serve as a plug-and-play solution, enhancing the performance of existing models.

Figures

Figures reproduced from arXiv: 2506.08955 by the authors.

Figure 1
Figure 1. Existing pseudo-labeling frameworks. Compared with previous methods ((a) and (b)), our SEE framework (c) is a unified [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Framework of SEE for ISCOS tasks. (a) In pseudo-label generation (PLG), SEE uses SAM to generate pseudo-labels [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Segmentation masks with different augmented images. We inversely transform the masks to keep consistent with the [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: Multi-density prompt extraction, including points, box, [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Details of HGFG. ΦN is feature grouping with N prototypes, whose broadcast process is simplified for space limitation. C C C Encoder Decoder HGFG D4 D3 D2 D1 JA JA JA E4 E0 E1 E2 E3 R R R R Xi ASPP i ' Y 2 Fi 3 Fi 4 Fi 1 Fi 2 Fi 3 Fi 4 Fi 1 Fi 1 ˆ Fi 2 Fi ˆ 3 Fi ˆ 4 Fi…
Figure 6
Figure 6. Figure 6: Framework of our segmenter with HGFG. Note that [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: Visualized results for COS tasks with point supervision, including COD, PIS, and TOD. [PITH_FULL_IMAGE:figures/full_fig_p010_7.png]
Figure 8
Figure 8. Figure 8: Visualized results for COS tasks with 1/8 semi-supervision, including COD, PIS, and TOD. [PITH_FULL_IMAGE:figures/full_fig_p012_8.png]
Figure 9
Figure 9. Figure 9: Analysis of point number C. TABLE X: Ablation studies of pseudo-label storage and supervision. Metrics Stor. #1 Stor. #2 Stor. #3 Stor. #4 Sup. #1 Sup. #2 Sup. #3 Ours M ↓ 0.037 0.036 0.038 0.038 0.039 0.038 0.038 0.036 Fβ ↑ 0.724 0.723 0.716 0.718 0.708 0.715 0.723 0.…
Figure 11
Figure 11. Figure 11: Visualization of feature similarity among group prototypes and grid features for [PITH_FULL_IMAGE:figures/full_fig_p015_11.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

66 extracted references · 57 canonical work pages

  1. [1]

    Concealed object detection,

    D.-P. Fan, G.-P. Ji, M.-M. Cheng, and L. Shao, “Concealed object detection,”IEEE Trans. Pattern Anal. Mach. Intell., 2021. 1, 11

    work page 2021

  2. [2]

    Image threshold segmentation based on glle histogram,

    X. Wang, L. Deng, and G. Xu, “Image threshold segmentation based on glle histogram,” inCPSCom. IEEE, 2019, pp. 410–415. 1

    work page 2019

  3. [3]

    Strategic preys make acute predators: Enhancing camouflaged object detectors by generating camouflaged objects,

    C. He, K. Li, Y . Zhang, Y . Zhang, Z. Guo, and X. Li, “Strategic preys make acute predators: Enhancing camouflaged object detectors by generating camouflaged objects,”ICLR, 2024. 1, 3

    work page 2024

  4. [4]

    A survey of camouflaged object detection and beyond,

    F. Xiao, S. Hu, Y . Shen, and C. He, “A survey of camouflaged object detection and beyond,”CAAI AIR, 2024. 1

    work page 2024

  5. [5]

    Concealed object segmentation with hierarchical coherence modeling,

    F. Xiao, P. Zhang, and C. He, “Concealed object segmentation with hierarchical coherence modeling,” inCAAI ICAI, 2023, pp. 16–27. 1

    work page 2023

  6. [6]

    Zoomnext: A unified collaborative pyramid network for camouflaged object detection,

    Y . Pang, X. Zhao, T.-Z. Xiang, L. Zhang, and H. Lu, “Zoomnext: A unified collaborative pyramid network for camouflaged object detection,” IEEE Trans. Pattern Anal. Mach. Intell., 2024. 1

    work page 2024

  7. [7]

    Detecting camouflaged object in frequency domain,

    Y . Zhong, B. Li, L. Tang, and S. Kuang, “Detecting camouflaged object in frequency domain,” inCVPR, 2022, pp. 4504–4513. 1

    work page 2022

  8. [8]

    Camouflaged object detection with feature decomposition and edge reconstruction,

    C. He, K. Li, Y . Zhang, L. Tang, and Y . Zhang, “Camouflaged object detection with feature decomposition and edge reconstruction,” inCVPR, 2023, pp. 22 046–22 055. 1, 8, 15, 16

    work page 2023

Show all 66 references
  1. [9]

    Diffusion models in low-level vision: A survey,

    C. He, Y . Shen, C. Fang, F. Xiao, and L. Tang, “Diffusion models in low-level vision: A survey,”IEEE Trans. Pattern Anal. Mach. Intell.,

  2. [10]

    Spider: A unified framework for context-dependent concept segmentation,

    X. Zhao, Y . Pang, W. Ji, B. Sheng, and J. Zuo, “Spider: A unified framework for context-dependent concept segmentation,” inICML. 1

  3. [11]

    Weakly-supervised concealed object segmentation with sam-based pseudo labeling and multi-scale feature grouping,

    C. He, K. Li, Y . Zhang, G. Xu, and L. Tang, “Weakly-supervised concealed object segmentation with sam-based pseudo labeling and multi-scale feature grouping,”NeurIPS, 2024. 1, 3, 5, 9, 11, 16

    2024

  4. [12]

    Camoteacher: Dual- rotation consistency learning for semi-supervised camouflaged object detection,

    X. Lai, Z. Yang, J. Hu, S. Zhang, and L. Cao, “Camoteacher: Dual- rotation consistency learning for semi-supervised camouflaged object detection,”arXiv preprint arXiv:2408.08050, 2024. 1, 3

    2024 arXiv

  5. [13]

    Real-world image dehazing with coherence-based label generator and cooperative unfolding network,

    C. Fang, C. He, F. Xiao, Y . Zhang, L. Tang, Y . Zhang, K. Li, and X. Li, “Real-world image dehazing with coherence-based label generator and cooperative unfolding network,”NeurIPS, 2025. 1, 12

    2025

  6. [14]

    Reti-diff: Illumination degradation image restoration with retinex-based latent diffusion model,

    C. He, C. Fang, Y . Zhang, T. Ye, K. Li, L. Tang, Z. Guo, X. Li, and S. Farsiu, “Reti-diff: Illumination degradation image restoration with retinex-based latent diffusion model,”ICLR, 2025. 1

    2025

  7. [15]

    Unfoldir: Rethinking deep unfolding network in illumination degrada- tion image restoration,

    C. He, R. Zhang, F. Xiao, C. Fang, L. Tang, Y . Zhang, and S. Farsiu, “Unfoldir: Rethinking deep unfolding network in illumination degrada- tion image restoration,”arXiv preprint arXiv:2505.06683, 2025. 1

    2025 arXiv

  8. [16]

    Segment anything,

    A. Kirillov, E. Mintun, N. Ravi, and H. Mao, “Segment anything,”arXiv preprint arXiv:2304.02643, 2023. 3, 4, 9, 10, 11, 13, 16

    2023 arXiv

  9. [17]

    Can sam segment anything? sam meets camou- flaged object detection,

    L. Tang and H. Xiao, “Can sam segment anything? sam meets camou- flaged object detection,”arXiv preprint arXiv:2304.04709, 2023. 3

    2023 arXiv

  10. [18]

    Sam struggles in concealed scenes

    G.-P. Ji, D.-P. Fan, P. Xu, M.-M. Cheng, and B. Zhou, “Sam struggles in concealed scenes”,”arXiv preprint arXiv:2304.06022, 2023. 3

    2023 arXiv

  11. [19]

    Segment anything is not always perfect: An investigation of sam on different real-world applications,

    W. Ji, J. Li, Q. Bi, W. Li, and L. Cheng, “Segment anything is not always perfect: An investigation of sam on different real-world applications,” arXiv preprint arXiv:2304.05750, 2023. 3

    2023 arXiv

  12. [20]

    Segment anything model for medical image analysis: an experimental study,

    M. A. Mazurowski, H. Dong, H. Gu, J. Yang, N. Konz, and Y . Zhang, “Segment anything model for medical image analysis: an experimental study,”arXiv preprint arXiv:2304.10517, 2023. 3

    2023 arXiv

  13. [21]

    Hqg-net: Unpaired medical image enhancement with high-quality guidance,

    C. He, K. Li, G. Xu, J. Yan, L. Tang, and Y . Zhang, “Hqg-net: Unpaired medical image enhancement with high-quality guidance,”IEEE Trans. Neural Networks Learn. Syst., 2023. 3

    2023

  14. [22]

    Sam-adapter: Adapting segment anything in underperformed scenes,

    T. Chen, L. Zhu, C. Deng, R. Cao, Y . Wang, and S. Zhang, “Sam-adapter: Adapting segment anything in underperformed scenes,” inICCV, 2023, pp. 3367–3375. 3, 9, 10, 11, 13, 16

    2023

  15. [23]

    Relax image-specific prompt requirement in sam: A single generic prompt for segmenting camouflaged objects,

    J. Hu, J. Lin, and S. Gong, “Relax image-specific prompt requirement in sam: A single generic prompt for segmenting camouflaged objects,” inAAAI, vol. 38, no. 11, 2024, pp. 12 511–12 518. 3, 9, 11, 16

    2024

  16. [24]

    Pranet: Parallel reverse attention network for polyp segmentation,

    D.-P. Fan, G.-P. Ji, and T. Zhou, “Pranet: Parallel reverse attention network for polyp segmentation,” inMICCAI, 2020, pp. 263–273. 3

    2020

  17. [25]

    Camouflaged object detection,

    D.-P. Fan, G.-P. Ji, G. Sun, M.-M. Cheng, and J. Shen, “Camouflaged object detection,” inCVPR, 2020, pp. 2777–2787. 3, 8, 10

    2020

  18. [26]

    Rich context aggregation with reflection prior for glass surface detection,

    J. Lin and Z. He, “Rich context aggregation with reflection prior for glass surface detection,” inCVPR, 2021, pp. 13 415–13 424. 3, 11 IEEE TPAMI 18

    2021

  19. [27]

    Referring camouflaged object detection,

    X. Zhang, B. Yin, Z. Lin, Q. Hou, D.-P. Fan, and M.-M. Cheng, “Referring camouflaged object detection,”IEEE Trans. Pattern Anal. Mach. Intell., 2025. 3

    2025

  20. [28]

    Camoformer: Masked separable attention for camouflaged object detection,

    B. Yin, X. Zhang, D.-P. Fan, S. Jiao, M.-M. Cheng, L. Van Gool, and Q. Hou, “Camoformer: Masked separable attention for camouflaged object detection,”IEEE Trans. Pattern Anal. Mach. Intell., 2024. 3

    2024

  21. [29]

    Run: Reversible unfolding network for concealed object segmentation,

    C. He, R. Zhang, F. Xiao, C. Fang, L. Tang, and Y . Zhang, “Run: Reversible unfolding network for concealed object segmentation,”arXiv preprint arXiv:2501.18783, 2025. 3, 10

    2025 arXiv

  22. [30]

    Un- locking attributes’ contribution to successful camouflage: A combined textual and visual analysis strategy,

    H. Zhang, Y . Lyu, Q. Yu, H. Liu, H. Ma, D. Yuan, and Y . Yang, “Un- locking attributes’ contribution to successful camouflage: A combined textual and visual analysis strategy,” inECCV. Springer, 2024, pp. 315–331. 3

    2024

  23. [31]

    Leveraging hallucinations to reduce manual prompt dependency in promptable segmentation,

    J. Hu, J. Lin, J. Yan, and S. Gong, “Leveraging hallucinations to reduce manual prompt dependency in promptable segmentation,”NeurIPS, vol. 37, pp. 107 171–107 197, 2025. 3

    2025

  24. [32]

    Weakly-supervised camouflaged object detection with scribble,

    R. He, Q. Dong, and J. Lin, “Weakly-supervised camouflaged object detection with scribble,”AAAI, 2023. 3, 9, 11, 15, 16, 17

    2023

  25. [33]

    Deep residual learning for image recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” inCVPR, 2016, pp. 770–778. 8

    2016

  26. [34]

    Degradation-resistant unfolding network for heterogeneous image fusion,

    C. He, K. Li, and Y . Zhang, “Degradation-resistant unfolding network for heterogeneous image fusion,” inICCV, 2023, pp. 611–621. 8

    2023

  27. [35]

    Object-centric learning with slot attention,

    F. Locatello, D. Weissenborn, T. Unterthiner, A. Mahendran, and G. Heigold, “Object-centric learning with slot attention,”NeurIPS, vol. 33, pp. 11 525–11 538, 2020. 8, 12

    2020

  28. [36]

    Weakly-supervised salient object detection via scribble,

    J. Zhang, X. Yu, A. Li, P. Song, and B. Liu, “Weakly-supervised salient object detection via scribble,” inCVPR, 2020, pp. 12 546–12 555. 9, 17

    2020

  29. [37]

    Structure-consistent weakly supervised salient object detection with local saliency coherence,

    S. Yu, B. Zhang, J. Xiao, and E. G. Lim, “Structure-consistent weakly supervised salient object detection with local saliency coherence,” in AAAI, vol. 35, 2021, pp. 3234–3242. 9, 16

    2021

  30. [38]

    Tree energy loss: Towards sparsely annotated semantic segmentation,

    Z. Liang, T. Wang, X. Zhang, J. Sun, and J. Shen, “Tree energy loss: Towards sparsely annotated semantic segmentation,” inCVPR, 2022, pp. 16 907–16 916. 9, 15, 16, 17

    2022

  31. [39]

    Towards automated polyp segmentation using weakly-and semi-supervised learning and deformable transformers,

    G. Ren, M. Lazarou, J. Yuan, and T. Stathaki, “Towards automated polyp segmentation using weakly-and semi-supervised learning and deformable transformers,” inCVPR, 2023, pp. 4355–4364. 11, 13

    2023

  32. [40]

    Pseudo-label guided contrastive learning for semi- supervised medical image segmentation,

    H. Basak and Z. Yin, “Pseudo-label guided contrastive learning for semi- supervised medical image segmentation,” inCVPR, 2023, pp. 19 786– 19 797. 11, 13, 16

    2023

  33. [41]

    Saliency as pseudo-pixel supervision for weakly and semi-supervised semantic segmentation,

    M. Lee, S. Lee, J. Lee, and H. Shim, “Saliency as pseudo-pixel supervision for weakly and semi-supervised semantic segmentation,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 45, no. 10, pp. 12 341– 12 357, 2023. 11, 12, 13, 15, 16, 17

    2023

  34. [42]

    Unsupervised and semi-supervised co-salient object detection via segmentation fre- quency statistics,

    S. Chakraborty, S. Naha, M. Bastan, D. Samaraset al., “Unsupervised and semi-supervised co-salient object detection via segmentation fre- quency statistics,” inWACV, 2024, pp. 332–342. 11, 13, 15, 16, 17

    2024

  35. [43]

    Randaugment: Practical automated data augmentation with a reduced search space,

    E. D. Cubuk, B. Zoph, J. Shlens, and Q. V . Le, “Randaugment: Practical automated data augmentation with a reduced search space,” inCVPRW, 2020, pp. 702–703. 13, 14

    2020

  36. [44]

    Denseaspp for semantic segmentation in street scenes,

    M. Yang, K. Yu, C. Zhang, Z. Li, and K. Yang, “Denseaspp for semantic segmentation in street scenes,” inCVPR, 2018, pp. 3684–3692. 8

    2018

  37. [45]

    Psanet: Point-wise spatial attention network for scene parsing,

    H. Zhao, Y . Zhang, S. Liu, and J. Shi, “Psanet: Point-wise spatial attention network for scene parsing,” inECCV, 2018, pp. 267–283. 8

    2018

  38. [46]

    Eca-net: Efficient channel attention for deep convolutional neural networks,

    Q. Wang, B. Wu, P. Zhu, P. Li, W. Zuo, and Q. Hu, “Eca-net: Efficient channel attention for deep convolutional neural networks,” inCVPR, 2020, pp. 11 534–11 542. 8

    2020

  39. [47]

    Image super-resolution using very deep residual channel attention networks,

    Y . Zhang, K. Li, and K. Li, “Image super-resolution using very deep residual channel attention networks,” inECCV, 2018, pp. 286–301. 9

    2018

  40. [48]

    Imagenet: A large- scale hierarchical image database,

    J. Deng, W. Dong, R. Socher, L.-J. Li, and K. Li, “Imagenet: A large- scale hierarchical image database,” inCVPR, 2009, pp. 248–255. 10

    2009

  41. [49]

    Integrating extra modality helps segmentor find camouflaged objects well,

    C. Fang, C. He, L. Tang, Y . Zhang, C. Zhu, Y . Shen, C. Chen, G. Xu, and X. Li, “Integrating extra modality helps segmentor find camouflaged objects well,”arXiv preprint arXiv:2502.14471, 2025. 10

    2025 arXiv

  42. [50]

    How to evaluate foreground maps?

    R. Margolin, L. Zelnik-Manor, and A. Tal, “How to evaluate foreground maps?” inCVPR, 2014, pp. 248–255. 10

    2014

  43. [51]

    Cognitive vision inspired object segmentation metric and loss function,

    D.-P. Fan, G.-P. Ji, X. Qin, and M.-M. Cheng, “Cognitive vision inspired object segmentation metric and loss function,”Scientia Sinica Informationis, vol. 6, no. 6, 2021. 10

    2021

  44. [52]

    Structure-measure: A new way to evaluate foreground maps,

    D.-P. Fan, M.-M. Cheng, Y . Liu, and T. Li, “Structure-measure: A new way to evaluate foreground maps,” inICCV, 2017, pp. 4548–4557. 10

    2017

  45. [53]

    Animal camouflage analysis: Chameleon database,

    P. Skurowski, H. Abdulameer, and J. Błaszczyk, “Animal camouflage analysis: Chameleon database,”Unpublished manuscript, p. 7, 2018. 11

    2018

  46. [54]

    Anabranch network for camouflaged object segmentation,

    T.-N. Le, T. V . Nguyen, Z. Nie, M.-T. Tran, and A. Sugimoto, “Anabranch network for camouflaged object segmentation,”Comput. Vis. Image Underst., vol. 184, pp. 45–56, 2019. 11

    2019

  47. [55]

    Simultaneously localize, segment and rank the camouflaged objects,

    Y . Lv, J. Zhang, Y . Dai, A. Li, B. Liu, N. Barnes, and D.-P. Fan, “Simultaneously localize, segment and rank the camouflaged objects,” inCVPR, 2021, pp. 11 591–11 601. 11

    2021

  48. [56]

    Automated polyp detection in colonoscopy videos using shape and context information,

    N. Tajbakhsh, S. R. Gurudu, and J. Liang, “Automated polyp detection in colonoscopy videos using shape and context information,”IEEE Trans. Med. Imaging, vol. 35, no. 2, pp. 630–644, 2015. 11

    2015

  49. [57]

    Toward embedded detection of polyps in wce images for early diagnosis,

    J. Silva, A. Histace, O. Romain, and X. Dray, “Toward embedded detection of polyps in wce images for early diagnosis,”Int. J. Comput. Assist. Radiol. Surg., vol. 9, pp. 283–293, 2014. 11

    2014

  50. [58]

    Kvasir-seg: A segmented polyp dataset,

    D. Jha, P. H. Smedsrud, M. A. Riegler, and P. Halvorsen, “Kvasir-seg: A segmented polyp dataset,” inMMM, 2020, pp. 451–462. 11

    2020

  51. [59]

    Don’t hit me! glass detection in real-world scenes,

    H. Mei, X. Yang, Y . Wang, Y . Liu, and S. He, “Don’t hit me! glass detection in real-world scenes,” inCVPR, 2020, pp. 3687–3696. 11

    2020

  52. [60]

    Weakly-supervised salient object detection using point supervison,

    S. Gao, W. Zhang, Y . Wang, Q. Guo, C. Zhang, Y . He, and W. Zhang, “Weakly-supervised salient object detection using point supervison,” in AAAI, vol. 36, 2022, pp. 670–678. 11, 17

    2022

  53. [61]

    Sam 2: Segment anything in images and videos,

    N. Ravi, V . Gabeur, Y .-T. Hu, R. Hu, C. Ryaliet al., “Sam 2: Segment anything in images and videos,”arXiv preprint arXiv:2408.00714, 2024. 12

    2024 arXiv

  54. [62]

    Exploring figure-ground assignment mechanism in perceptual organization,

    W. Zhai, Y . Cao, and J. Zhang, “Exploring figure-ground assignment mechanism in perceptual organization,” inNeurIPS, vol. 35, 2022. 15, 16

    2022

  55. [63]

    Pyramid grafting network for one-stage high resolution saliency detection,

    C. Xie, C. Xia, M. Ma, Z. Zhao, X. Chen, and J. Li, “Pyramid grafting network for one-stage high resolution saliency detection,” inCVPR, 2022, pp. 11 717–11 726. 15, 17

    2022

  56. [64]

    Pixels, regions, and objects: Multiple enhancement for salient object detection,

    Y . Wang, R. Wang, X. Fan, T. Wang, and X. He, “Pixels, regions, and objects: Multiple enhancement for salient object detection,” inCVPR, 2023, pp. 10 031–10 040. 15, 17

    2023

  57. [65]

    Learning to detect salient objects with image-level supervision,

    L. Wang, H. Lu, and Y . Wang, “Learning to detect salient objects with image-level supervision,” inCVPR, 2017, pp. 136–145. 16 Chunming Hereceived the B.S. degree from Nanjing University of Posts and Telecommunications, China, and the M.E. degree from Tsinghua University, Chin...

    2017

  58. [2022]

    Outstanding Member of the Editorial Board Award

    His current research interests are machine learning and computer vision, including weakly supervised learning and 3-D scene understanding. Ziyun Yangreceived his Ph.D. degree from Department of Biomedical Engineering at Duke University. He received his B.S. degree in Automatio...

    2019

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.