Pith. sign in

REVIEW 4 major objections 6 minor 51 references

ST-SAM: SAM-Driven Self-Training Framework for Semi-Supervised Camouflaged Object Detection

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read ST-SAM claims a single self-trained network with entropy-filtered pseudo-labels and SAM-guided correction reaches fully supervised COD performance from 1% labels.

desk verdict ST-SAM is the first to put a frozen SAM prompt-correction loop inside self-training for camouflaged object detection, and the reported gains are large and consistent — but the entropy filter's treatment of all-background pseudo-labels is a real unexamined flaw that undercuts the error-correction claim. read the letter →

arxiv 2507.23307 v1 pith:D36TEPDN submitted 2025-07-31 cs.CV

Xihang Hu , Fuming Sun , Jiazhe Liu , Feilong Xu , Xiaoli Zhang This is my paper
classification cs.CV
keywords Semi-supervisedSelf-TrainingCamouflagedobjectdetectionSegmentAnythingModelPseudo-labelfilteringEntropy-baseddynamicDomainprompt-guidedmutualcorrectionAnnotationefficiency
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

ST-SAM is a semi-supervised camouflaged object detection framework built around a single trainable network, and its central claim is that with only 1% pixel-level labels this network can reach performance comparable to fully supervised COD models. The paper argues that the usual failure mode of self-training, where noisy pseudo-labels accumulate into collapsed predictions, can be prevented by filtering pseudo-labels through an entropy-based uncertainty measure and then feeding the surviving labels as box-and-point prompts to a frozen SAM whose masks are fused back with the model's own predictions. On four COD benchmarks (CAMO, CHAMELEON, COD10K, NC4K) the paper reports that this loop beats the existing teacher-student SSCOD methods at the same annotation budgets and matches or exceeds several fully supervised baselines. A reader should care because the claim, if true, makes dense annotation largely unnecessary for camouflage detection and suggests a general recipe for bootstrapping vision models with almost no labels.

What carries the argument

The object that carries the argument is the pair of strategies EDF and DPC. EDF computes local entropy $E_{local}$ and global entropy $E_{global}$ of each initial pseudo-label, defines the uncertainty metric $u_\alpha$ as the fraction of pixels with $E_{local} > 0.5 E_{global}$, retains samples with $u_\alpha < \tau_\alpha$, and builds the entropy-weighted pseudo-label $P^E_i = P^I_i (0.5 + 0.5(1-E_{local})^k)$. DPC computes box prompts from the minimum bounding rectangles of the largest mask components and point prompts from geometric centers, with an axial search to move centers that fall outside the mask back inside, then obtains $P^S_i = \mathrm{SAM}(Image, Prompt_{P-B})$ and sets $P^C_i = \mathrm{fuse}(P^E_i, P^S_i)$. The identity doing the work is the equal-proportion fusion: the paper reports that intersection loses true positives, union adds false positives, and 1:1 fusion best suppresses error, while the dynamic low-to-high entropy expansion prevents early collapse.

What would settle it

Run a 1%-label COD10K training in which the SAM prompt is derived from a randomly selected contour of $P^E_i$ instead of the full filtered mask. If the weighted F-measure stays near 0.713, then the entropy filter is doing the work and the prompt design is not load-bearing; if it drops sharply, the DPC prompt-to-mask correction is responsible for the reported gain.

Watch

Extended reading notes

Core claim

The discovery the paper is trying to establish is that error accumulation in self-training for COD can be broken without any teacher-student architecture, task-specific segmentation loss, or SAM fine-tuning. The mechanism is a mutual-correction loop: the student network generates initial pseudo-labels; EDF keeps only samples whose local-vs-global entropy ratio is below $\tau_\alpha=0.3$ and reweights each kept pixel by its local entropy; DPC then converts the weighted masks into a hybrid prompt (minimum-bounding-rectangle box plus point prompts with axial-search safe centers) and asks a frozen SAM to segment the image from those prompts; the SAM mask and the entropy-weighted pseudo-label are fused in equal proportion to form the new training label. The paper's reported numbers at 1% labels, for example weighted F-measure on COD10K rising from 0.517 for the supervised-only baseline to 0.713 for the full ST-SAM, are offered as evidence that the filter and the SAM correction each contribute and combine additively.

Load-bearing premise

The load-bearing premise is that entropy-based confidence reliably identifies which pseudo-labels and pixels are trustworthy in camouflaged scenes, and that prompting a frozen SAM with boxes and points derived from those labels yields masks accurate enough to correct errors rather than compound them.

Editorial extensions

If this is right

  • At 1% labeled data, the framework is reported to outperform the two existing SSCOD methods even when those methods use up to 20% labels, and to match or exceed several fully supervised detectors on all four benchmarks.
  • The framework does not depend on a specific COD network or loss: swapping to a lighter network or to BCE-IoU lowers performance slightly but preserves a large margin over the SSCOD baseline.
  • The easy-to-hard expansion schedule is necessary for the result: adding all qualified pseudo-labels at once drops COD10K weighted F-measure to 0.621, while the epoch-dynamic schedule reaches 0.713.
  • Because SAM is used frozen with box and point prompts, the method avoids the computational cost of fine-tuning a foundation model, which is part of the paper's claim of practical scalability.

Reading between the lines

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

  • If the same loop transfers to other tasks with entangled foreground-background structure, such as medical lesion segmentation or industrial defect detection, entropy-filtered self-training plus prompt-corrected foundation-model masks could become a generic low-annotation recipe; the paper only demonstrates this for COD.
  • The paper does not report variance across random draws of the 1% labeled subset, so the robustness of the headline 1% result to which images happen to be labeled is an open question.
  • Equal-proportion fusion suppresses errors only when the two label sources fail independently; if SAM and the student share the same hallucination, the fused label will not remove it. A natural extension would be to measure per-pixel agreement between $P^E_i$ and $P^S_i$ and study whether final errors cluster where both agree.
  • The reported 'concise' claim is about architecture, not compute: the pipeline runs a heavyweight SAM (vit-h) over the unlabeled set during expansion, so total inference cost is higher than a single forward pass of the COD network.
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

4 major / 6 minor

Summary. The paper proposes ST-SAM, a semi-supervised camouflaged object detection (SSCOD) framework that replaces the teacher-student architectures used by prior methods with a single segmentation network trained via self-training. A network pretrained on a small labeled set generates pseudo-labels on unlabeled data; Entropy-based Dynamic Filtering (EDF) retains low-uncertainty samples, applies pixel-level entropy weighting, and progressively expands the training set; Domain Prompt-guided Mutual Correction (DPC) converts filtered pseudo-labels into box/point prompts for a frozen SAM and fuses SAM outputs with the filtered pseudo-labels. Experiments on CAMO, CHAMELEON, COD10K, and NC4K at 1%, 5%, 10%, and 20% labeled data report large improvements over CamoTeacher and SSCOD, and ablations show that each component contributes. Scalability tests indicate that replacement of the backbone or loss still yields strong results.

Significance. The main value of this work is empirical: if the results hold, ST-SAM is a simple, annotation-efficient alternative to existing teacher-student SSCOD methods, and it provides a concrete way to inject foundation-model knowledge into self-training without fine-tuning. The reported gains are large and consistent across four datasets, and the component ablations in Tables 2-5 support the claim that both EDF and DPC matter. The framework is model-agnostic, and the scalability experiment is a useful contribution. However, the headline 1%-labeled result is not yet backed by a fully reproducible protocol: the dynamic expansion schedule is unspecified, no multiple-seed statistics are reported, and the proposed error-correction mechanism is not directly validated. These gaps are within the scope of a revision.

major comments (4)
  1. [Section 3.2, Eq. (5); Section 3.3, Eqs. (7)-(9)] The retention rule in Eq. (5) cannot reject the most important failure mode under 1% supervision: a complete false negative. If the network outputs a background-only mask, then E_local = E_global = 0 and u_alpha = 0, so the empty pseudo-label is retained. Eq. (6) leaves it empty, and because DPC builds box and point prompts from mask contours (Eqs. (7)-(8)), no prompt is generated and SAM cannot repair the mask in Eq. (9). The paper does not report object-level recall of retained pseudo-labels, the fraction of retained empty masks, or an IoU comparison between P_E and P_S, so the central claim that EDF and DPC filter and correct pseudo-label errors is unverified for false negatives.
  2. [Section 3.2, dynamic expansion] The dynamic expansion schedule is specified only verbally: 'Assuming the current training set size is x, the top x low-entropy samples ... are selected ... As the training epochs increase, x will gradually increase.' No formula, algorithm, or concrete schedule is given, and the supplementary material promised in Section 4.3 is not included in this submission. This is load-bearing because Table 3 shows that the epoch-dynamic schedule is the best variant (Fw 0.713 vs 0.621 for one-shot expansion), so without the schedule the reported gains cannot be reproduced.
  3. [Section 4.1 and Table 1] All experiments, including the 1%-labeled comparison that anchors the headline claim, are reported from a single random split with no error bars, multiple seeds, or significance tests. At 1% of the 4,040-image training set there are only about 40 labeled images, so the differences between ST-SAM and the SSCOD baselines may depend substantially on which images are sampled. The paper should report mean and standard deviation over at least three seeds, especially for the 1% and 5% settings.
  4. [Section 3.3, Eq. (9); Table 4] The 'mutual correction' mechanism is asserted but not directly validated. The paper does not report how often the SAM output P_S has higher IoU with the true mask than the entropy-weighted P_E, how the two outputs disagree, or whether disagreements are resolved in the correct direction. Given the paper's own citations [22, 36] that SAM struggles in concealed scenes, an agreement analysis is necessary before one can conclude that DPC injects domain knowledge rather than additional noise; at minimum, per-sample IoU statistics for P_E, P_S, and the fused P_C should be reported.
minor comments (6)
  1. [Section 4.2 and Table 1] The abstract's claim of 'even matching fully supervised methods' is stronger than Table 1 supports: on CAMO and COD10K, fully supervised methods such as DSAM still lead on E_xi, S_alpha, and F_beta, while ST-SAM is competitive mainly on MAE and on NC4K; 'competitive with' would be a more accurate phrasing.
  2. [Section 4.3, Table 3] The claim that learning low-entropy samples first brings further improvement over random selection is not supported by the numbers: Random Select and L-to-H Entropy differ by 0.001 in E_xi and 0.000-0.003 in Fw on COD10K and NC4K. This wording should be softened or the comparison repeated over multiple seeds.
  3. [Table 1] The results are non-monotonic in the amount of labeled data for Ours-5% on COD10K (Fw 0.709 vs 0.713 at 1%, and F_beta 0.716 vs 0.737 at 1%), but this is not discussed.
  4. [Eq. (10)] The Dice loss denominator appears to contain a typo: 'Σ+Target_p' should presumably be Σ Target_p, giving 1 - 2*Σ(Pred_p*Target_p)+S over Σ Pred_p + Σ Target_p + S.
  5. [Section 3.4] The terms w_bce and w_iou in the structural loss L_S are not defined, and the fusion operation fuse(.,.) in Eq. (9) is described only as 'equal proportion'; these should be specified for reproducibility.
  6. [Section 1] The introduction contains grammatical errors, including 'we proposes' and 'we achieves', which should be corrected before publication.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: ST-SAM's entropy filtering and SAM-prompt correction are heuristic pipeline components, not derivations that reduce to their inputs; the empty-mask failure mode is a real limitation but not circularity.

full rationale

The paper's contribution is an empirical self-training pipeline rather than a mathematical derivation. Pseudo-labels are generated by the same network being trained, which is standard self-training bootstrapping, but the two load-bearing components are an entropy-based retention gate (Eqs. 4-6) and a frozen SAM prompted from those pseudo-labels (Eqs. 7-9). Neither component is defined in terms of the final reported metric, and the hyperparameters (tau_alpha=0.3, k=1, alpha=4, beta=2) are fixed constants rather than quantities fitted inside the equations. SAM is an external model trained on SA-1B and is not fine-tuned on COD, so the SAM output P_S is not the same object as the entropy-weighted pseudo-label P_E by construction. The only self-citation, Ref. [20] (PRNet), is used as a drop-in network in the scalability ablation (Table 5); this supports a secondary claim and is not load-bearing. A genuine limitation, but not a circularity, is the false-negative failure mode: an all-background pseudo-label has E_local=E_global=0, so u_alpha=0 in Eq. (4) and Eq. (5) retains it; Eq. (6) leaves it empty; Eqs. (7)-(8) then generate no box or point prompts, so Eq. (9) cannot repair the missed object. The paper does not report object-level recall of retained pseudo-labels, IoU between P_E and P_S, or the fraction of empty retained masks, so this robustness assumption is unverified. That is a correctness concern, not an equivalence between output and input by construction.

Assumptions & free parameters 7 free parameters · 4 assumptions · 0 invented entities

The central claim depends on several hand-set hyperparameters (tau_alpha, omega, k, alpha, beta, lambda_ual, fusion ratio) and on two substantive assumptions: that entropy reliably measures pseudo-label quality in camouflaged scenes, and that a frozen SAM can be prompt-steered to correct COD pseudo-labels. No new entities are invented.

free parameters (7)
  • tau_alpha (uncertainty threshold) = 0.3
    Used in Eq. 5 to decide which pseudo-labeled samples are retained; chosen by hand and affects the filtering strength.
  • omega (local entropy window size) = 7
    Window size for local mean and entropy computation in Eqs. 1-2; set without reported tuning.
  • k (entropy weight decay coefficient) = 1
    Controls the strength of pixel-level entropy weighting in Eq. 6.
  • alpha (Dice loss weight) = 4
    Loss weight in Eq. 12 for the Dice term; hand-picked.
  • beta (uncertainty-aware loss weight) = 2
    Loss weight in Eq. 12 for the uncertainty-aware term; hand-picked.
  • lambda_ual (dynamic weighting coefficient) = decreases with learning rate (not specified)
    Weighting coefficient in Eq. 11 for the uncertainty-aware loss; the schedule is not reported.
  • fusion ratio for P_E and P_S = 0.5
    Equal-proportion fusion in Eq. 9; fixed, but ablation shows it outperforms intersection and union.
assumptions (4)
  • ad hoc to paper Entropy-based uncertainty metric correlates with pseudo-label quality
    The retention rule (Eq. 5) and pixel weighting (Eq. 6) assume that low local entropy relative to global entropy indicates more reliable predictions in camouflaged scenes; the paper offers no theoretical justification.
  • domain assumption Frozen SAM can be steered by box and point prompts to segment camouflaged objects
    The DPC strategy relies on SAM (vit-h) producing useful masks when prompted with noisy pseudo-label boxes and points, despite prior evidence that SAM struggles in concealed scenes (refs [22], [36]).
  • domain assumption The random 1% labeled split is representative of the training distribution
    All experiments use a single random split of labeled vs unlabeled data; no multiple-seed analysis is reported, so results may depend on which samples are labeled.
  • domain assumption The classic U-Net COD network (from ref [45]) is a suitable basis for COD
    The framework adopts a standard encoder-decoder with ConvNeXt backbone, Frequency Transformer, and ASPP module; this is a reasonable baseline architecture but its suitability is assumed.

how reviews work

0 comments
Cite this review

Pith. "Pith review of ST-SAM: SAM-Driven Self-Training Framework for Semi-Supervised Camouflaged Object Detection." pith.science (2026). https://pith.science/paper/D36TEPDN

@misc{pith2026250723307,
  author       = {Pith},
  title        = {Pith review of: ST-SAM: SAM-Driven Self-Training Framework for Semi-Supervised Camouflaged Object Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/D36TEPDN}},
  note         = {Machine review of arXiv:2507.23307}
}
read the original abstract

Semi-supervised Camouflaged Object Detection (SSCOD) aims to reduce reliance on costly pixel-level annotations by leveraging limited annotated data and abundant unlabeled data. However, existing SSCOD methods based on Teacher-Student frameworks suffer from severe prediction bias and error propagation under scarce supervision, while their multi-network architectures incur high computational overhead and limited scalability. To overcome these limitations, we propose ST-SAM, a highly annotation-efficient yet concise framework that breaks away from conventional SSCOD constraints. Specifically, ST-SAM employs Self-Training strategy that dynamically filters and expands high-confidence pseudo-labels to enhance a single-model architecture, thereby fundamentally circumventing inter-model prediction bias. Furthermore, by transforming pseudo-labels into hybrid prompts containing domain-specific knowledge, ST-SAM effectively harnesses the Segment Anything Model's potential for specialized tasks to mitigate error accumulation in self-training. Experiments on COD benchmark datasets demonstrate that ST-SAM achieves state-of-the-art performance with only 1\% labeled data, outperforming existing SSCOD methods and even matching fully supervised methods. Remarkably, ST-SAM requires training only a single network, without relying on specific models or loss functions. This work establishes a new paradigm for annotation-efficient SSCOD. Codes will be available at https://github.com/hu-xh/ST-SAM.

Figures

Figures reproduced from arXiv: 2507.23307 by the authors.

Figure 1
Figure 1. Performance comparison of semi-supervised COD [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. The comparison of typical SSCOD frameworks: (a) [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. The overall flowchart of ST-SAM. Tr: The Net is preliminarily trained using labeled data [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (2 more)
Figure 5
Figure 5. Figure 5: The overall architecture of COD network, consisting [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]
Figure 6
Figure 6. Figure 6: Qualitative comparison results between ST-SAM and SOTA COD models. [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

51 extracted references · 34 canonical work pages

  1. [22]

    Segment Anything

    Ge-Peng Ji, Deng-Ping Fan, Peng Xu, Bowen Zhou, Ming-Ming Cheng, and Luc Van Gool. 2023. SAM struggles in concealed scenes - empirical study on "Segment Anything". Sci. China Inf. Sci. 66, 12 (2023). https://doi.org/10.1007/ S11432-023-3881-X

    work page 2023

  2. [36]

    Lv Tang, Haoke Xiao, and Bo Li. 2023. Can SAM Segment Anything? When SAM Meets Camouflaged Object Detection. CoRR abs/2304.04709 (2023). https: //doi.org/10.48550/ARXIV.2304.04709 arXiv:2304.04709

    arXiv doi:10.48550/arxiv.2304.04709 2023

  3. [1]

    Hemami, Francisco J

    Radhakrishna Achanta, Sheila S. Hemami, Francisco J. Estrada, and Sabine Süsstrunk. 2009. Frequency-tuned salient region detection. In 2009 IEEE Com- puter Society Conference on Computer Vision and Pattern Recognition (CVPR 2009), 20-25 June 2009, Miami, Florida, USA . IEEE Computer Society, 1597–1604. https://doi.org/10.1109/CVPR.2009.5206596

    work page arXiv 2009

  4. [2]

    Massih-Reza Amini, Vasilii Feofanov, Loïc Pauletto, Lies Hadjadj, Emilie Devijver, and Yury Maximov. 2025. Self-training: A survey. Neurocomputing 616 (2025), 128904. https://doi.org/10.1016/J.NEUCOM.2024.128904

    work page arXiv 2025

  5. [3]

    Huafeng Chen, Dian Shao, Guangqian Guo, and Shan Gao. 2024. Just a Hint: Point- Supervised Camouflaged Object Detection. In Computer Vision - ECCV 2024 - 18th European Conference, Milan, Italy, September 29-October 4, 2024, Proceedings, Part XXXV (Lecture Notes in Computer Science, Vol. 15093) , Ales Leonardis, Elisa Ricci, Stefan Roth, Olga Russakovsky, ...

    work page doi:10.1007/978-3-031-72761-0_19 2024

  6. [4]

    Huafeng Chen, Pengxu Wei, Guangqian Guo, and Shan Gao. 2024. SAM-COD: SAM-Guided Unified Framework for Weakly-Supervised Camouflaged Object Detection. In Computer Vision - ECCV 2024 - 18th European Conference, Milan, Italy, September 29-October 4, 2024, Proceedings, Part XXXV (Lecture Notes in Computer Science, Vol. 15093) , Ales Leonardis, Elisa Ricci, S...

    work page doi:10.1007/978-3-031-72761-0_18 2024

  7. [5]

    Liang-Chieh Chen, George Papandreou, Florian Schroff, and Hartwig Adam

  8. [6]

    Tianrun Chen, Lanyun Zhu, Chaotao Ding, Runlong Cao, Yan Wang, Shangzhan Zhang, Zejian Li, Lingyun Sun, Ying Zang, and Papa Mao. 2023. SAM-Adapter: Adapting Segment Anything in Underperformed Scenes. In IEEE/CVF Interna- tional Conference on Computer Vision, ICCV 2023 - Workshops, Paris, France, October 2-6, 2023. IEEE, 3359–3367. https://doi.org/10.1109/...

    arXiv 2023

Show all 51 references
  1. [7]

    Deng-Ping Fan, Ming-Ming Cheng, Yun Liu, Tao Li, and Ali Borji. 2017. Structure- Measure: A New Way to Evaluate Foreground Maps. In IEEE International Con- ference on Computer Vision, ICCV 2017, Venice, Italy, October 22-29, 2017 . IEEE Computer Society, 4558–4567. https://doi...

    2017 doi

  2. [8]

    Deng-Ping Fan, Cheng Gong, Yang Cao, Bo Ren, Ming-Ming Cheng, and Ali Borji

  3. [9]

    Deng-Ping Fan, Ge-Peng Ji, Peng Xu, Ming-Ming Cheng, Christos Sakaridis, and Luc Van Gool. 2023. Advances in deep concealed scene understanding. Vis. Intell. 1, 1 (2023). https://doi.org/10.1007/S44267-023-00019-6

    2023 doi

  4. [10]

    Deng-Ping Fan, Ge-Peng Ji, Guolei Sun, Ming-Ming Cheng, Jianbing Shen, and Ling Shao. 2020. Camouflaged object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition . 2777–2787

    2020

  5. [12]

    Chao Hao, Zitong Yu, Xin Liu, Jun Xu, Huanjing Yue, and Jing-Yu Yang. 2025. A Simple Yet Effective Network Based on Vision Transformer for Camouflaged Object and Salient Object Detection. IEEE Trans. Image Process. 34 (2025), 608–622. https://doi.org/10.1109/TIP.2025.3528347

    2025

  6. [13]

    Chunming He, Kai Li, Yachao Zhang, Guoxia Xu, Longxiang Tang, Yulun Zhang, Zhenhua Guo, and Xiu Li. 2023. Weakly-supervised concealed object segmenta- tion with sam-based pseudo labeling and multi-scale feature grouping. Advances in Neural Information Processing Systems 36 (20...

    2023

  7. [14]

    Chunming He, Kai Li, Yachao Zhang, Yulun Zhang, Chenyu You, Zhenhua Guo, Xiu Li, Martin Danelljan, and Fisher Yu. 2024. Strategic Preys Make Acute Predators: Enhancing Camouflaged Object Detectors by Generating Camouflaged Objects. In The Twelfth International Conference on Le...

    2024

  8. [15]

    Ruozhen He, Qihua Dong, Jiaying Lin, and Rynson W. H. Lau. 2023. Weakly- Supervised Camouflaged Object Detection with Scribble Annotations. In Thirty- Seventh AAAI Conference on Artificial Intelligence, AAAI 2023, Thirty-Fifth Con- ference on Innovative Applications of Artific...

    2023

  9. [17]

    Jie Hu, Chen Chen, Liujuan Cao, Shengchuan Zhang, Annan Shu, Guannan Jiang, and Rongrong Ji. 2023. Pseudo-label Alignment for Semi-supervised Instance Segmentation. In IEEE/CVF International Conference on Computer Vision, ICCV 2023, Paris, France, October 1-6, 2023 . IEEE, 162...

    2023

  10. [18]

    Jian Hu, Jiayi Lin, Shaogang Gong, and Weitong Cai. 2024. Relax Image-Specific Prompt Requirement in SAM: A Single Generic Prompt for Segmenting Cam- ouflaged Objects. In Thirty-Eighth AAAI Conference on Artificial Intelligence, AAAI 2024, Thirty-Sixth Conference on Innovative...

    2024

  11. [19]

    Jian Hu, Jiayi Lin, Junchi Yan, and Shaogang Gong. 2024. Leveraging Hallucinations to Reduce Manual Prompt Dependency in Promptable Seg- mentation. In Advances in Neural Information Processing Systems 38: An- nual Conference on Neural Information Processing Systems 2024, NeurI...

    2024

  12. [20]

    Xihang Hu, Xiaoli Zhang, Fasheng Wang, Jing Sun, and Fuming Sun. 2024. Ef- ficient Camouflaged Object Detection Network Based on Global Localization Perception and Local Guidance Refinement. IEEE Trans. Circuits Syst. Video Technol. 34, 7 (2024), 5452–5465. https://doi.org/10....

    2024

  13. [21]

    Wenjun Hui, Zhenfeng Zhu, Shuai Zheng, and Yao Zhao. 2024. Endow SAM with Keen Eyes: Temporal-Spatial Prompt Learning for Video Camouflaged Object Detection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, W A, USA, June 16-22, 2024. IEEE...

    2024

  14. [23]

    Kingma and Jimmy Ba

    Diederik P. Kingma and Jimmy Ba. 2015. Adam: A Method for Stochastic Opti- mization. In 3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings , Yoshua Bengio and Yann LeCun (Eds.). http://arxiv.org/...

    2015 arXiv

  15. [24]

    Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al

  16. [25]

    Xunfa Lai, Zhiyu Yang, Jie Hu, Shengchuan Zhang, Liujuan Cao, Guannan Jiang, Zhiyu Wang, Songan Zhang, and Rongrong Ji. 2024. CamoTeacher: Dual-Rotation Consistency Learning for Semi-supervised Camouflaged Object Detection. In Computer Vision - ECCV 2024 - 18th European Confer...

    2024

  17. [26]

    Nguyen, Zhongliang Nie, Minh-Triet Tran, and Akihiro Sugimoto

    Trung-Nghia Le, Tam V. Nguyen, Zhongliang Nie, Minh-Triet Tran, and Akihiro Sugimoto. 2019. Anabranch network for camouflaged object segmentation. Com- put. Vis. Image Underst. 184 (2019), 45–56. https://doi.org/10.1016/J.CVIU.2019. 04.006

    2019 doi

  18. [28]

    Yunqiu Lv, Jing Zhang, Yuchao Dai, Aixuan Li, Bowen Liu, Nick Barnes, and Deng- Ping Fan. 2021. Simultaneously Localize, Segment and Rank the Camouflaged Objects. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021 . Computer Vis...

    2021

  19. [29]

    Ran Margolin, Lihi Zelnik-Manor, and Ayellet Tal. 2014. How to Evaluate Fore- ground Maps. In 2014 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2014, Columbus, OH, USA, June 23-28, 2014. IEEE Computer Society, 248–255. https://doi.org/10.1109/CVPR.2014.39 M...

    2014 doi

  20. [30]

    Muhammad Nawfal Meeran, Gokul Adethya T, and Bhanu Pratyush Mantha

  21. [32]

    Yuzhen Niu, Lifen Yang, Rui Xu, Yuezhou Li, and Yuzhong Chen. 2024. MiNet: Weakly-Supervised Camouflaged Object Detection through Mutual Interaction between Region and Edge Cues. In Proceedings of the 32nd ACM International Conference on Multimedia, MM 2024, Melbourne, VIC, Au...

    2024

  22. [33]

    Youwei Pang, Xiaoqi Zhao, Tian-Zhu Xiang, Lihe Zhang, and Huchuan Lu

  23. [34]

    Youwei Pang, Xiaoqi Zhao, Jiaming Zuo, Lihe Zhang, and Huchuan Lu. 2025. Open-Vocabulary Camouflaged Object Segmentation. InComputer Vision – ECCV 2024, Aleš Leonardis, Elisa Ricci, Stefan Roth, Olga Russakovsky, Torsten Sattler, and Gül Varol (Eds.). Springer Nature Switzerla...

    2025

  24. [35]

    Przemysław Skurowski, Hassan Abdulameer, Jakub Błaszczyk, Tomasz Depta, Adam Kornacki, and Przemysław Kozieł. 2018. Animal camouflage analysis: Chameleon database. Unpublished manuscript 2, 6 (2018), 7

    2018

  25. [37]

    Zhe Tang, Jing Tang, Dengpeng Zou, Junyi Rao, and Fang Qi. 2024. Two guidance joint network based on coarse map and edge map for camouflaged object detec- tion. Appl. Intell. 54, 15-16 (2024), 7531–7544. https://doi.org/10.1007/S10489- 024-05559-Y

    2024 doi

  26. [38]

    Bingyang Wang, Tanlin Li, Jiannan Wu, Yi Jiang, Huchuan Lu, and You He. 2023. A Simple Baseline for Open-World Tracking via Self-training. In Proceedings of the 31st ACM International Conference on Multimedia, MM 2023, Ottawa, ON, Canada, 29 October 2023- 3 November 2023, Abdu...

    2023 doi

  27. [39]

    Song Wu, Xiaoyu Wei, Xinyue Chen, Yazhou Ren, Jing He, and Xiaorong Pu. 2024. Cross-View Mutual Learning for Semi-Supervised Medical Image Segmentation. In Proceedings of the 32nd ACM International Conference on Multimedia, MM 2024, Melbourne, VIC, Australia, 28 October 2024 -...

    2024

  28. [40]

    Chenxi Xie, Changqun Xia, Tianshu Yu, and Jia Li. 2023. Frequency Represen- tation Integration for Camouflaged Object Detection. In Proceedings of the 31st ACM International Conference on Multimedia, MM 2023, Ottawa, ON, Canada, 29 October 2023- 3 November 2023, Abdulmotaleb E...

    2023

  29. [41]

    Lihe Yang, Zhen Zhao, and Hengshuang Zhao. 2025. UniMatch V2: Pushing the Limit of Semi-Supervised Semantic Segmentation. IEEE Trans. Pattern Anal. Mach. Intell. 47, 4 (2025), 3031–3048. https://doi.org/10.1109/TPAMI.2025.3528453

    2025

  30. [42]

    Lihe Yang, Wei Zhuo, Lei Qi, Yinghuan Shi, and Yang Gao. 2022. ST++: Make Self- trainingWork Better for Semi-supervised Semantic Segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022 . IEEE, 4258–4267...

    2022

  31. [43]

    Mingzhao Yang, Shangchao Su, Bin Li, and Xiangyang Xue. 2024. Exploring One- Shot Semi-supervised Federated Learning with Pre-trained Diffusion Models. In Thirty-Eighth AAAI Conference on Artificial Intelligence, AAAI 2024, Thirty- Sixth Conference on Innovative Applications o...

    2024

  32. [44]

    Zhenni Yu, Xiaoqin Zhang, Li Zhao, Yi Bin, and Guobao Xiao. 2024. Exploring Deeper! Segment Anything Model with Depth Perception for Camouflaged Object Detection. In Proceedings of the 32nd ACM International Conference on Multimedia, MM 2024, Melbourne, VIC, Australia, 28 Octo...

    2024

  33. [45]

    Jin Zhang, Ruiheng Zhang, Yanjiao Shi, Zhe Cao, Nian Liu, and Fahad Shahbaz Khan. 2024. Learning Camouflaged Object Detection from Noisy Pseudo Label. In Computer Vision - ECCV 2024 - 18th European Conference, Milan, Italy, Sep- tember 29-October 4, 2024, Proceedings, Part I (...

    2024 doi

  34. [46]

    Shiwei Zhang, Wei Ke, Shuai Liu, Xiaopeng Hong, and Tong Zhang. 2024. Boost- ing Semi-supervised Crowd Counting with Scale-based Active Learning. In Pro- ceedings of the 32nd ACM International Conference on Multimedia, MM 2024, Mel- bourne, VIC, Australia, 28 October 2024 - 1 ...

    2024

  35. [47]

    Xuying Zhang, Bowen Yin, Zheng Lin, Qibin Hou, Deng-Ping Fan, and Ming- Ming Cheng. 2025. Referring Camouflaged Object Detection. IEEE Transactions on Pattern Analysis and Machine Intelligence 47, 5 (2025), 3597–3610. https: //doi.org/10.1109/TPAMI.2025.3532440

    2025

  36. [48]

    Zhen Zhao, Lihe Yang, Sifan Long, Jimin Pi, Luping Zhou, and Jingdong Wang. 2023. Augmentation Matters: A Simple-Yet-Effective Approach to Semi- Supervised Semantic Segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canad...

    2023

  37. [49]

    Ke Zhou, Zhongwei Qiu, and Dongmei Fu. 2024. Multi-scale contrastive adaptor learning for segmenting anything in underperformed scenes. Neurocomputing 606 (2024), 128395. https://doi.org/10.1016/J.NEUCOM.2024.128395

    2024

  38. [50]

    Xiaofei Zhou, Zhicong Wu, and Runmin Cong. 2024. Decoupling and Integration Network for Camouflaged Object Detection. IEEE Trans. Multim. 26 (2024), 7114–7129. https://doi.org/10.1109/TMM.2024.3360710

    2024

  39. [2017]

    CoRR abs/1706.05587 (2017)

    Rethinking Atrous Convolution for Semantic Image Segmentation. CoRR abs/1706.05587 (2017). arXiv:1706.05587 http://arxiv.org/abs/1706.05587

    2017 arXiv

  40. [2018]

    In Proceedings of the Twenty-Seventh International Joint Conference on Artificial Intelligence, IJCAI 2018, July 13-19, 2018, Stockholm, Sweden , Jérôme Lang (Ed.)

    Enhanced-alignment Measure for Binary Foreground Map Evaluation. In Proceedings of the Twenty-Seventh International Joint Conference on Artificial Intelligence, IJCAI 2018, July 13-19, 2018, Stockholm, Sweden , Jérôme Lang (Ed.). ijcai.org, 698–704. https://doi.org/10.24963/IJ...

    2018 doi

  41. [2022]

    In IEEE/CVF Conference on Computer Vision and Pattern Recog- nition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022

    Zoom In and Out: A Mixed-scale Triplet Network for Camouflaged Ob- ject Detection. In IEEE/CVF Conference on Computer Vision and Pattern Recog- nition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022 . IEEE, 2150–2160. https://doi.org/10.1109/CVPR52688.2022.00220

    2022

  42. [2023]

    In Proceedings of the IEEE/CVF international conference on computer vision

    Segment anything. In Proceedings of the IEEE/CVF international conference on computer vision. 4015–4026

  43. [2024]

    In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024 - Workshops, Seattle, W A, USA, June 17-18, 2024

    SAM-PM: Enhancing Video Camouflaged Object Detection using Spatio- Temporal Attention. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024 - Workshops, Seattle, W A, USA, June 17-18, 2024 . IEEE, 1857–1866. https://doi.org/10.1109/CVPRW63382.2024.00192

    2024

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.