Pith. sign in

REVIEW 4 major objections 4 minor 70 references

Stepwise Decomposition and Dual-stream Focus: A Novel Approach for Training-free Camouflaged Object Segmentation

T0 review · 4 major / 4 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 The paper introduces RDVP-MSD, a training-free pipeline that claims state-of-the-art camouflaged-object segmentation from a single generic prompt by decomposing captions into foreground/background text and constraining point prompts to a…

desk verdict Decent engineering result in training-free COS, but the SOTA claim leans on test-set-tuned hyperparameters and missing code; worth reviewing, not worth trusting at face value. read the letter →

arxiv 2506.06818 v3 pith:HSAPCTVU submitted 2025-06-07 cs.CV

Chao Yin , Hao Li , Kequan Yang , Jide Li , Pinpin Zhu , Xiaoqiang Li This is my paper
classification cs.CV
keywords camouflagedobjectsegmentationtraining-freetest-timeadaptationpromptablemultimodalchainofthoughtvisualpromptingSegmentAnythingModelregion-constraineddual-streamself-consistencymaskselection
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

The paper introduces RDVP-MSD, a training-free test-time pipeline that segments camouflaged objects from a single generic prompt such as "camouflaged object." It claims to resolve two failure modes of prior promptable methods: ambiguous text prompts generated from captions, and visual prompts whose background points come from far away and semantically unrelated regions. To fix the text side, it decomposes captions through a four-step multimodal chain of thought into foreground and background phrases and words; to fix the visual side, it computes separate foreground and background heatmaps inside the predicted bounding box, samples only points above a 0.9 confidence threshold, and feeds them to a segmentation foundation model in coarse-to-fine stages. On COD10K, CAMO, and CHAMELEON the method reports state-of-the-art task-generic-prompt segmentation accuracy, surpassing GenSAM and ProMaC while running 3.6× and 8.5× faster, all without any training. If correct, automatic image-level prompting could substitute for manual point and box prompts in camouflaged-scene segmentation.

What carries the argument

The load-bearing machinery is a pair of modules. MSD-CoT (Multimodal Stepwise Decomposition Chain of Thought) is a four-step MLLM reasoning chain—caption generation, phrase disentanglement, keyword identification, coarse location—that turns the image into separate foreground and background text prompts plus a bounding box. RDVP (Region-constrained Dual-stream Visual Prompting) is the second module: it generates independent foreground and background heatmaps with the VLM, constrains both to the predicted box, and selects point prompts at pixels whose normalized heat value exceeds 0.9, as written in Equations (6) and (7). The coarse-to-fine Text-to-Mask Generator feeds phrase-level prompts into the VFM to get a preliminary mask, refines the box, then feeds word-level prompts; Self-Consistency Mask Selection runs three parallel repetitions and picks the mask closest to the mean. The threshold-0.9 point selection inside the box is the mechanism that is supposed to stop SAM-family VFMs from leaking onto unrelated background regions.

What would settle it

Run RDVP-MSD on the COD10K and CAMO test splits with the language branch replaced by oracle text prompts, such as ground-truth class labels from the datasets, and compare with the full method; the accuracy gap isolates how much error MSD-CoT injects. Independently, sweep the 0.9 threshold in Equations (6) and (7) across 0.5 to 0.99 and measure $S_\alpha$, $F_\beta$, and $M$; if performance is flat across a wide threshold range, the claim that high-confidence in-box points are what matters weakens, whereas a sharp peak at 0.9 would confirm the paper's assumed heatmap calibration.

Watch

Extended reading notes

Core claim

RDVP-MSD claims that camouflaged object segmentation can be solved in a fully training-free, task-generic promptable mode by replacing global consensus heatmaps with region-constrained dual-stream visual prompting and replacing direct caption-to-prompt translation with stepwise multimodal decomposition. The text branch (MSD-CoT) generates a caption, disentangles it into phrase-level foreground and background descriptions, condenses those into word-level keywords, and asks the MLLM for a coarsely located bounding box. The visual branch (RDVP) then uses a frozen vision-language model to create two separate heatmaps, one for foreground text and one for background text, normalizes them, and selects point prompts only where the normalized response inside the bounding box is at least 0.9; the coarse mask from those prompts tightens the box, and word-level prompts repeat the procedure for fine-grained refinement. A self-consistency step runs three parallel mask candidates under synonym task-generic prompts and keeps the mask closest to the mean. The paper reports that this combination achieves the best reported accuracy in the task-generic prompt setting on COD10K, CAMO, and CHAMELEON, and that it surpasses weak-supervision baselines that use point or scribble labels.

Load-bearing premise

The load-bearing premise is that the vision-language model's normalized heatmaps, computed separately for the MLLM's foreground and background text phrases, place the true camouflaged object's pixels above the 0.9 threshold inside the predicted bounding box; if the language prompt is semantically wrong or the heatmap cannot localize the object, every downstream point prompt and mask inherits that error.

Editorial extensions

If this is right

  • Task-generic promptable segmentation of camouflaged scenes no longer requires manual points, boxes, or per-image language prompts; the generic phrase "camouflaged object" suffices.
  • Reported accuracy on COD10K, CAMO, and CHAMELEON exceeds prior task-generic promptable methods GenSAM and ProMaC, and matches or beats weakly supervised methods that use point or scribble supervision.
  • Inference is faster than the baselines, at about 19.85 seconds per image averaged across datasets, with a 3.6× speedup over GenSAM and an 8.5× speedup over ProMaC, and the three repetitions are independent so they can run in parallel.
  • The stepwise decomposition helps when holistic captions are semantically ambiguous, for example when a caption mentions "grass" and a prior pipeline mistakes grass for a foreground prompt.
  • Because the method is entirely training-free, it can be applied to new camouflaged scenes without collecting pixel labels or fine-tuning models.

Reading between the lines

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

  • Inference: if the core claim holds, the same two-stage design—decompose a task-level prompt into sub-object language, then constrain visual prompt sampling to the estimated region—should transfer to other tasks where generic prompts hide instance ambiguity, such as salient object detection in cluttered scenes or rare-category segmentation.
  • Inference: the 0.9 threshold in Equations (6) and (7) is a calibration assumption that the paper does not directly test; if a VLM's heatmap ranks the true camouflaged pixels below 0.9 inside the box, RDVP's point set loses the object, so an oracle-language experiment separating MLLM text errors from VLM heatmap errors would locate the bottleneck.
  • Inference: because the method uses frozen components, a natural extension is to sweep the confidence threshold across a range such as 0.5 to 0.99; a sharp performance peak at 0.9 would confirm the paper's assumed heatmap calibration, whereas flat performance would indicate the threshold is not the decisive factor.
  • Inference: the self-consistency selection step implies that stochastic MLLM outputs are a real source of error; a testable extension is to compare the final mask's quality against the best individual repetition to measure how much consensus voting recovers from noisy text prompts.
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 / 4 minor

Summary. The paper proposes RDVP-MSD, a training-free, task-generic promptable segmentation framework for camouflaged object segmentation (COS). The method composes three frozen models: LLaVA-1.5-13B for multimodal stepwise decomposition chain-of-thought (MSD-CoT) that generates hierarchical foreground/background text prompts and a coarse bounding box; Spatial CLIP for region-constrained dual-stream visual prompting (RDVP) that produces separate foreground and background heatmaps and samples high-confidence points inside the bounding box; and HQ-SAM for mask generation in a coarse-to-fine scheme. A self-consistency mask selection runs I parallel repetitions with different generic prompts and picks the mask closest to the mean. Experiments on COD10K, CAMO, and CHAMELEON report state-of-the-art results under the task-generic prompt setting, with claimed gains over GenSAM and ProMaC and 3.6x/8.5x speedups.

Significance. If the empirical claims hold, this is a useful contribution: it replaces manual instance-specific prompts with automated text and visual prompt generation, removes the iterative feedback loops of prior train-free methods, and demonstrates that a composition of frozen foundation models can reach competitive COS accuracy without training. The paper has clear strengths: explicit ablation of MSD-CoT, RDVP, and its two strategies; comparison across three standard benchmarks; and a concrete efficiency analysis. The main concerns are that the central SOTA claim rests on hyperparameters selected on the test benchmarks, that the thresholded heatmap equations are under-specified, and that the stochastic pipeline is evaluated with single-run numbers. These issues are fixable in revision and do not invalidate the overall approach, but they currently prevent the headline result from being read as a predictive, reproducible statement.

major comments (4)
  1. [§4.3, Table 4] The repetition number I is selected as 3 because it yields the best S_alpha and F_beta on COD10K-TEST and CAMO-TEST. This is model selection on exactly the test benchmarks used for the state-of-the-art claim in Table 1. The RDVP threshold of 0.9 in Eqs. (6)-(7) is never ablated or justified. The headline improvement is therefore an in-sample selection rather than a predictive statement. Please provide a validation-based procedure or report sensitivity of the final metrics to I and to the threshold; if the reported numbers are oracle choices, this must be stated explicitly.
  2. [§3.3.1, Eqs. (6)-(7)] The normalization function N(·) is left undefined. If it is a per-heatmap min-max normalization, the threshold 0.9 is relative and its effect is not comparable across images; if it is a global sigmoid or softmax, the threshold has a different meaning. Moreover, if no pixel inside the predicted bounding box satisfies the threshold, the sets p_f^i and p_b^i are empty, and the paper describes no fallback. The method needs a precise definition of N(·), an ablation or sensitivity analysis of the threshold, and a report of how often empty point sets occur and how VFM handles them.
  3. [§3.2.1 and §3.4] The MSD-CoT prompt templates are only given as examples, not as the exact strings used, and the sampling configuration of the MLLM (temperature, number of samples, seed) is not reported. Because Self-Consistency Mask Selection relies on stochastic outputs, the single-run numbers in Tables 1-4 are insufficient to assess reliability. Please provide the full prompts and inference settings, and report mean and standard deviation over multiple runs or, at minimum, a fixed-seed reproducibility statement.
  4. [§3.2.1, Coarse Location] The text says that for highly camouflaged objects an image-level bounding box is used instead of the MLLM-predicted box, but no criterion is given for detecting 'highly camouflaged' cases. Since the benefit of RDVP depends on the bounding box being reasonably tight, the absence of this criterion weakens the claimed spatial constraint. Please specify how this fallback is triggered and how often it occurs on the evaluation sets.
minor comments (4)
  1. [§3.4, Eq. (9)] The 'mean mask' in Eq. (9) is not defined; please state whether masks are averaged as binary maps, probability maps, or after some normalization.
  2. [§4.1] There are several wording and formatting issues: 'These include' in the evaluation metrics sentence should be followed by a colon; '1, 250images' should be '1,250 images'; and 'besides, using 3times' in the implementation details has a missing space.
  3. [§4.2] The qualitative comparison in Figure 4 would be stronger if it also showed failure cases or images where RDVP-MSD struggles, given that no failure analysis is reported in the text.
  4. [Abstract and Section 1] The statement 'The codes will be available at ...' is not verifiable in the manuscript; if code is not yet public, please mark it as anonymous or supplementary material for review, or provide an anonymized archive.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the method is a zero-shot composition of frozen external models, and the central SOTA claim is an empirical comparison against external baselines; the self-citations are not load-bearing.

full rationale

RDVP-MSD is a training-free pipeline that chains three frozen external models (LLaVA for text prompts, Spatial CLIP as VLM, HQ-SAM as VFM). None of the paper's equations defines an output in terms of the evaluation target or a fitted parameter that is renamed as a prediction. The foreground and background phrases, heatmaps, point threshold (Eqs. 6-7), and mask selection (Eq. 9) are generation heuristics; the final mask is one of the candidate SAM outputs, not a quantity that reduces to its own input. The SOTA claim in Table 1 is an external comparison on standard benchmarks, and the cited self-works [50,51] appear only as related-work references and do not support any load-bearing premise. The repetition count I=3 is chosen by test-set performance in Table 4 and the 0.9 threshold is not ablated; these are validity and overfitting concerns about the reported accuracy, but they do not make the derivation circular because the masks themselves are produced by frozen VLMs and SAM, not by optimizing those hyperparameters. No equation-level reduction or fitted-input-as-prediction step was found. Consequently, the circularity score is 0.

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

The paper's contribution is a prompt and point-selection recipe layered on three frozen foundation models. The central claim rests on those models behaving as expected on camouflaged inputs, which the paper does not independently test. There are no new physical or mathematical entities, but there are several hand-chosen hyperparameters that affect the reported numbers.

free parameters (4)
  • Heatmap confidence threshold = 0.9
    Equations (6)-(7) select only normalized heatmap values >= 0.9; no sensitivity analysis is reported.
  • Self-consistency repetition count I = 3
    Ablated on COD10K-TEST and CAMO-TEST; I=3 is reported best among 1 to 6, suggesting selection on the test sets used for the main claim.
  • Task-generic prompt set P_g = camouflaged object, camouflaged animal, camouflaged entity
    Manually chosen diverse synonyms; the self-consistency vote depends on their diversity.
  • MLLM query templates = not fully specified
    Queries Q_s, Q_p, Q_w, Q_bbox are paraphrased in Section 3.2 but exact templates are not given; results are sensitive to prompt wording.
assumptions (4)
  • domain assumption LLaVA-1.5-13B generates captions, disentangled phrases, keywords, and bounding boxes reliably enough for COS.
    The whole MSD-CoT chain depends on this; no failure-rate analysis is provided.
  • domain assumption Spatial CLIP heatmaps correlate with semantic foreground and background for camouflaged objects.
    RDVP point selection in Equations (6)-(7) trusts these heatmaps to separate object from background, the exact failure mode of camouflage.
  • domain assumption HQ-SAM produces correct masks from the sampled points and box.
    SAM is a black-box component; the method assumes its inductive biases transfer to camouflaged scenes.
  • domain assumption COS benchmarks and metrics are a valid measure of camouflaged object segmentation.
    Claims rely on S_alpha, F_beta, M, E_phi from prior literature; no discussion of metric limitations.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Stepwise Decomposition and Dual-stream Focus: A Novel Approach for Training-free Camouflaged Object Segmentation." pith.science (2026). https://pith.science/paper/HSAPCTVU

@misc{pith2026250606818,
  author       = {Pith},
  title        = {Pith review of: Stepwise Decomposition and Dual-stream Focus: A Novel Approach for Training-free Camouflaged Object Segmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HSAPCTVU}},
  note         = {Machine review of arXiv:2506.06818}
}
read the original abstract

While promptable segmentation (\textit{e.g.}, SAM) has shown promise for various segmentation tasks, it still requires manual visual prompts for each object to be segmented. In contrast, task-generic promptable segmentation aims to reduce the need for such detailed prompts by employing only a task-generic prompt to guide segmentation across all test samples. However, when applied to Camouflaged Object Segmentation (COS), current methods still face two critical issues: 1) \textit{\textbf{semantic ambiguity in getting instance-specific text prompts}}, which arises from insufficient discriminative cues in holistic captions, leading to foreground-background confusion; 2) \textit{\textbf{semantic discrepancy combined with spatial separation in getting instance-specific visual prompts}}, which results from global background sampling far from object boundaries with low feature correlation, causing SAM to segment irrelevant regions. To address the issues above, we propose \textbf{RDVP-MSD}, a novel training-free test-time adaptation framework that synergizes \textbf{R}egion-constrained \textbf{D}ual-stream \textbf{V}isual \textbf{P}rompting (RDVP) via \textbf{M}ultimodal \textbf{S}tepwise \textbf{D}ecomposition Chain of Thought (MSD-CoT). MSD-CoT progressively disentangles image captions to eliminate semantic ambiguity, while RDVP injects spatial constraints into visual prompting and independently samples visual prompts for foreground and background points, effectively mitigating semantic discrepancy and spatial separation. Without requiring any training or supervision, RDVP-MSD achieves a state-of-the-art segmentation result on multiple COS benchmarks and delivers a faster inference speed than previous methods, demonstrating significantly improved accuracy and efficiency. The codes will be available at \href{https://github.com/ycyinchao/RDVP-MSD}{https://github.com/ycyinchao/RDVP-MSD}

Figures

Figures reproduced from arXiv: 2506.06818 by the authors.

Figure 1
Figure 1. Superior Performance and Efficiency: The proposed [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Motivation of the proposed RDVP-MSD. (a) Text Prompts: (1) Existing methods directly extract category cues from [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Overview of the proposed RDVP-MSD framework. It comprises two core components: (1) Multimodal Stepwise [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (2 more)
Figure 3
Figure 3. Figure 3: In the coarse stage, phrase-level text prompts generate initial instance-specific visual prompts via RDVP, resulting in a preliminary segmentation mask 𝑀 ′ 𝑖 and a refined bounding box 𝐵𝑖 . Subsequently, in the fine-grained stage, the refined bounding box 𝐵𝑖 and the mo…
Figure 4
Figure 4. Figure 4: Qualitative comparison of the proposed RDVP-MSD with two main task-generic promptable segmentation methods. [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

70 extracted references · 63 canonical work pages

  1. [1]

    Radhakrishna Achanta, Sheila Hemami, Francisco Estrada, and Sabine Susstrunk

    work page

  2. [2]

    Huafeng Chen, Dian Shao, Guangqian Guo, and Shan Gao. 2024. Just a Hint: Point-Supervised Camouflaged Object Detection. InECCV. 332–348

    work page 2024

  3. [3]

    Huafeng Chen, Pengxu Wei, Guangqian Guo, and Shan Gao. 2024. SAM-COD+: SAM-guided Unified Framework for Weakly-Supervised Camouflaged Object Detection.IEEE Transactions on Circuits and Systems for Video Technology(2024)

    work page 2024

  4. [4]

    Tianrun Chen, Lanyun Zhu, Chaotao Deng, Runlong Cao, Yan Wang, Shangzhan Zhang, Zejian Li, Lingyun Sun, Ying Zang, and Papa Mao. 2023. Sam-adapter: Adapting segment anything in underperformed scenes. InICCV. 3367–3375

    work page 2023

  5. [5]

    Wenqing Chen, Weicheng Wang, Zhixuan Chu, Kui Ren, Zibin Zheng, and Zhichao Lu. 2024. Self-Para-Consistency: Improving Reasoning Tasks at Low Cost for Large Language Models. InACL. 14162–14167

    work page 2024

  6. [6]

    Deng-Ping Fan, Ming-Ming Cheng, Yun Liu, Tao Li, and Ali Borji. 2017. Structure- Measure: A New Way to Evaluate Foreground Maps. InICCV. 4558–4567

    work page 2017

  7. [7]

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

  8. [8]

    Deng-Ping Fan, Ge-Peng Ji, Guolei Sun, Ming-Ming Cheng, Jianbing Shen, and Ling Shao. 2020. Camouflaged Object Detection. InCVPR. 2774–2784

    work page 2020

Show all 70 references
  1. [9]

    Yuanbin Fu, Jie Ying, Houlei Lv, and Xiaojie Guo. 2024. Semi-supervised Camou- flaged Object Detection from Noisy Data. InACM MM. 4766–4775

    2024

  2. [10]

    Chunming He, Kai Li, Yachao Zhang, Longxiang Tang, Yulun Zhang, Zhenhua Guo, and Xiu Li. 2023. Camouflaged Object Detection with Feature Decomposition and Edge Reconstruction. InCVPR. 22046–22055

    2023

  3. [11]

    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. InNeurIPS, Vol. 36. 30726–30737

    2023

  4. [12]

    Ruozhen He, Qihua Dong, Jiaying Lin, and Rynson WH Lau. 2023. Weakly- supervised camouflaged object detection with scribble annotations. InAAAI, Vol. 37. 781–789

    2023

  5. [13]

    Jian Hu, Jiayi Lin, Shaogang Gong, and Weitong Cai. 2024. Relax Image-Specific Prompt Requirement in SAM: A Single Generic Prompt for Segmenting Camou- flaged Objects. InAAAI, Vol. 38. 12511–12518

    2024

  6. [14]

    Jian Hu, Jiayi Lin, Junchi Yan, and Shaogang Gong. 2024. Leveraging halluci- nations to reduce manual prompt dependency in promptable segmentation. In NeurIPS, Vol. 37. 107171–107197

    2024

  7. [15]

    Ge-Peng Ji, Deng-Ping Fan, Yu-Cheng Chou, Dengxin Dai, Alexander Liniger, and Luc Van Gool. 2023. Deep gradient learning for efficient camouflaged object detection.Machine Intelligence Research20, 1 (2023), 92–108

    2023

  8. [16]

    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”.Science China Information Sciences66, 12 (2023), 226101

    2023

  9. [17]

    Wei Ji, Jingjing Li, Qi Bi, Tingwei Liu, Wenbo Li, and Li Cheng. 2024. Correction to: Segment Anything Is Not Always Perfect: An Investigation of SAM on Different Real-world Applications.Machine Intelligence Research21, 6 (2024), 1215–1215

    2024

  10. [18]

    Lei Ke, Mingqiao Ye, Martin Danelljan, Yifan Liu, Yu-Wing Tai, Chi-Keung Tang, and Fisher Yu. 2023. Segment Anything in High Quality. InNeurIPS, Vol. 36. 29914–29934

    2023

  11. [19]

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

  12. [20]

    Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. 2022. Large language models are zero-shot reasoners. InNeurIPS, Vol. 35. 22199–22213

    2022

  13. [21]

    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 ECCV. 438–455

    2024

  14. [22]

    Nguyen, Zhongliang Nie, Minh-Triet Tran, and Aki- hiro Sugimoto

    Trung-Nghia Le, Tam V. Nguyen, Zhongliang Nie, Minh-Triet Tran, and Aki- hiro Sugimoto. 2019. Anabranch network for camouflaged object segmentation. Computer Vision and Image Understanding184 (2019), 45–56

    2019

  15. [23]

    Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. 2023. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. InICML. 19730–19742

    2023

  16. [24]

    Yi Li, Hualiang Wang, Yiqun Duan, Jiheng Zhang, and Xiaomeng Li. 2025. A closer look at the explainability of Contrastive language-image pre-training. Pattern Recognition162 (2025), 111409

    2025

  17. [25]

    Zhiyuan Liang, Tiancai Wang, Xiangyu Zhang, Jian Sun, and Jianbing Shen. 2022. Tree energy loss: Towards sparsely annotated semantic segmentation. InCVPR. 16907–16916

    2022

  18. [26]

    Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. 2024. Improved baselines with visual instruction tuning. InCVPR. 26296–26306

    2024

  19. [27]

    Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. 2023. Visual instruc- tion tuning. InNeurIPS, Vol. 36. 34892–34916

    2023

  20. [28]

    Keshun Liu, Aihua Li, Sen Yang, Changlong Wang, and Yuhua Zhang. 2025. Multi- scale attention and boundary-aware network for military camouflaged object detection using unmanned aerial vehicles.Signal, Image and Video Processing19, 1 (2025), 184

    2025

  21. [29]

    Xinran Liu, Lin Qi, Yuxuan Song, and Qi Wen. 2024. Depth awakens: A depth- perceptual attention fusion network for RGB-D camouflaged object detection. Image and Vision Computing143 (2024), 104924

    2024

  22. [30]

    Yiwen Liu, Xiaoyu Zhang, Jinchao Zhu, and Panlong Tan. 2025. Improving un- derwater camouflage object segmentation with dual-decoder attention network. The Journal of Supercomputing81, 1 (2025), 1–21

    2025

  23. [31]

    Chancharik Mitra, Brandon Huang, Trevor Darrell, and Roei Herzig. 2024. Com- positional chain-of-thought prompting for large multimodal models. InCVPR. 14420–14431

    2024

  24. [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. InACM MM. 6316–6325

    2024

  25. [33]

    OpenAI. 2024. GPT-4V: Enhancing GPT-4 for Visual Processing. https://www. openai.com Accessed: 2024-05-20

    2024

  26. [34]

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. 2021. Learning transferable visual models from natural language supervision. InICML. 8748–8763

    2021

  27. [35]

    Tianhe Ren, Shilong Liu, Ailing Zeng, Jing Lin, Kunchang Li, He Cao, Jiayu Chen, Xinyu Huang, Yukang Chen, Feng Yan, et al. 2024. Grounded sam: Assembling open-world models for diverse visual tasks.arXiv preprint arXiv:2401.14159 (2024)

    2024 arXiv

  28. [36]

    Przemysław Skurowski, Hassan Abdulameer, Jakub Błaszczyk, Tomasz Depta, Adam Kornacki, and Przemysław Kozieł. 2018. Animal camouflage analysis: Chameleon database.Unpublished manuscript2, 6 (2018), 7

    2018

  29. [37]

    Yujia Sun, Shuo Wang, Chenglizhao Chen, and Tian-Zhu Xiang. 2022. Boundary- Guided Camouflaged Object Detection. InIJCAI. 1335–1341

    2022

  30. [38]

    Yanguang Sun, Chunyan Xu, Jian Yang, Hanyu Xuan, and Lei Luo. 2024. Frequency-spatial entanglement learning for camouflaged object detection. In ECCV. 343–360

    2024

  31. [39]

    Lv Tang, Peng-Tao Jiang, Zhi-Hao Shen, Hao Zhang, Jin-Wei Chen, and Bo Li

  32. [40]

    Thi Thu Hang Truong and Trung Kien Tran. 2024. A style transfer-based aug- mentation approach for detecting military camouflaged object.JMST’s Section on Computer Science and Control Engineering.CSCE8 (2024), 44–54

    2024

  33. [41]

    Lei Wang, Yi Hu, Jiabang He, Xing Xu, Ning Liu, Hui Liu, and Heng Tao Shen

  34. [42]

    Liqiong Wang, Jinyu Yang, Yanfu Zhang, Fangyi Wang, and Feng Zheng. 2024. Depth-aware concealed crop detection in dense agricultural scenes. InCVPR. 17201–17211

    2024

  35. [43]

    Qingwang Wang, Xin Qu, Liyao Zhou, Pengcheng Jin, Chengbiao Fu, and Tao Shen. 2024. Edge-Guided Pixel Level Connected Component Assisted Camou- flaged Object Detection. InICIP. 4021–4027

    2024

  36. [44]

    Tingran Wang, Zaiyang Yu, Jianwei Fang, Jinlong Xie, Feng Yang, Huang Zhang, Liping Zhang, Minghua Du, Lusi Li, and Xin Ning. 2025. Multidimensional fusion of frequency and spatial domain information for enhanced camouflaged object detection.Information Fusion117 (2025), 102871

    2025

  37. [45]

    Chi, Sharan Narang, Aakanksha Chowdhery, and Denny Zhou

    Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc V Le, Ed H. Chi, Sharan Narang, Aakanksha Chowdhery, and Denny Zhou. 2023. Self-Consistency Improves Chain of Thought Reasoning in Language Models. InICLR

    2023

  38. [46]

    InAAAI, Vol

    T-sciq: Teaching multimodal chain-of-thought reasoning via large language model signals for science question answering. InAAAI, Vol. 38. 19162–19170

  39. [47]

    Zongwei Wu, Danda Pani Paudel, Deng-Ping Fan, Jingjing Wang, Shuo Wang, Cédric Demonceaux, Radu Timofte, and Luc Van Gool. 2023. Source-free Depth for Object Pop-out. InICCV. 1032–1042

    2023

  40. [48]

    Zongwei Wu, Jingjing Wang, Zhuyun Zhou, Zhaochong An, Qiuping Jiang, Cédric Demonceaux, Guolei Sun, and Radu Timofte. 2023. Object Segmentation by Mining Cross-Modal Semantics. InACM MM. 3455–3464

    2023

  41. [49]

    Chenxi Xie, Changqun Xia, Tianshu Yu, and Jia Li. 2023. Frequency representation integration for camouflaged object detection. InACM MM. 1789–1797. MM ’25, October 27–31, 2025, Dublin, Ireland Chao Yin et al

    2023

  42. [50]

    Chao Yin and Xiaoqiang Li. 2025. Dual region mutual enhancement network for camouflaged object detection.Image and Vision Computing158 (2025), 105526

    2025

  43. [51]

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. 2022. Chain-of-thought prompting elicits reasoning in large language models. InNeurIPS, Vol. 35. 24824–24837

    2022

  44. [52]

    Siyue Yu, Bingfeng Zhang, Jimin Xiao, and Eng Gee Lim. 2021. Structure- consistent weakly supervised salient object detection with local saliency co- herence. InAAAI, Vol. 35. 3234–3242

    2021

  45. [53]

    Zhenni Yu, Xiaoqin Zhang, Li Zhao, Yi Bin, and Guobao Xiao. 2024. Exploring Deeper! Segment Anything Model with Depth Perception for Camouflaged Object Detection. InACM MM. 4322–4330

    2024

  46. [54]

    Mingfeng Zha, Yunqiang Pei, Guoqing Wang, Tianyu Li, Yang Yang, Wenbin Qian, and Heng Tao Shen. 2024. Weakly-Supervised Mirror Detection via Scribble Annotations. InAAAI, Vol. 38. 6953–6961

    2024

  47. [55]

    Yuting Zhai, Zongmei Gao, Yang Zhou, Jian Li, Yuqi Zhang, and Yanlei Xu

  48. [56]

    Chao Yin, Kequan Yang, Jide Li, Xiaoqiang Li, and Yifan Wu. 2024. Camouflaged Object Detection via Complementary Information-Selected Network Based on Visual and Semantic Separation.IEEE Transactions on Industrial Informatics20, 11 (2024), 12871–12881

    2024

  49. [57]

    Jin Zhang, Ruiheng Zhang, Yanjiao Shi, Zhe Cao, Nian Liu, and Fahad Shahbaz Khan. 2024. Learning Camouflaged Object Detection from Noisy Pseudo Label. InECCV. 158–174

    2024

  50. [58]

    Xiaoqin Zhang, Zhenni Yu, Li Zhao, Deng-Ping Fan, and Guobao Xiao. 2025. COMPrompter: reconceptualized segment anything model with multiprompt network for camouflaged object detection.Science China Information Sciences68, 1 (2025), 112104

    2025

  51. [59]

    Zhuosheng Zhang, Aston Zhang, Mu Li, hai zhao, George Karypis, and Alex Smola. 2024. Multimodal Chain-of-Thought Reasoning in Language Models. Transactions on Machine Learning Research(2024)

    2024

  52. [60]

    Jianwei Zhao, Xin Li, Fan Yang, Qiang Zhai, Ao Luo, Zicheng Jiao, and Hong Cheng. 2024. Focusdiffuser: Perceiving local disparities for camouflaged object detection. InECCV. 181–198

    2024

  53. [61]

    Green fruit detection methods: Innovative application of camouflage object detection and multilevel feature mining.Computers and Electronics in Agriculture 225 (2024), 109356

    2024

  54. [62]

    Jing Zhang, Xin Yu, Aixuan Li, Peipei Song, Bowen Liu, and Yuchao Dai. 2020. Weakly-supervised salient object detection via scribble annotations. InCVPR. 12546–12555

    2020

  55. [63]

    Xueyan Zou, Zi-Yi Dou, Jianwei Yang, Zhe Gan, Linjie Li, Chunyuan Li, Xiyang Dai, Harkirat Behl, Jianfeng Wang, Lu Yuan, et al. 2023. Generalized decoding for pixel, image, and language. InCVPR. 15116–15127

    2023

  56. [64]

    Xueyan Zou, Jianwei Yang, Hao Zhang, Feng Li, Linjie Li, Jianfeng Wang, Lijuan Wang, Jianfeng Gao, and Yong Jae Lee. 2023. Segment everything everywhere all at once. InNeurIPS, Vol. 36. 19769–19782

    2023

  57. [67]

    Ke Zhou, Zhongwei Qiu, and Dongmei Fu. 2024. Multi-scale contrastive adaptor learning for segmenting anything in underperformed scenes.Neurocomputing 606 (2024), 128395

    2024

  58. [68]

    Zijian Zhu, Xiao Yang, Hang Su, and Shibao Zheng. 2025. CamoEnv: Transfer- able and environment-consistent adversarial camouflage in autonomous driving. Pattern Recognition Letters188 (2025), 95–102

    2025

  59. [2009]

    Frequency-tuned salient region detection. InCVPR. 1597–1604

  60. [2018]

    In IJCAI

    Enhanced-alignment Measure for Binary Foreground Map Evaluation. In IJCAI. 698–704

  61. [2023]

    Segment anything. InICCV. 4015–4026

  62. [2024]

    InACM MM

    Chain of visual perception: Harnessing multimodal large language models for zero-shot camouflaged object detection. InACM MM. 8805–8814

Pith tools

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