Pith. sign in

REVIEW 4 major objections 4 minor 62 references

ArgusCogito: Chain-of-Thought for Cross-Modal Synergy and Omnidirectional Reasoning in Camouflaged Object Segmentation

T0 review · 4 major / 4 minor · reviewed 2026-08-05 · deepseek-v4-flash

Pith's one-line read ArgusCogito claims a zero-shot VLM chain-of-thought—conjecture, focus, sculpting—can segment camouflaged objects at state-of-the-art accuracy without task-specific training.

desk verdict A plausible three-stage VLM+SAM2 pipeline, but the zero-shot SOTA claim is compromised by test-set tuning and contradicted by the paper's own tables. read the letter →

arxiv 2508.18050 v1 pith:CZE64HDE submitted 2025-08-25 cs.CV

classification cs.CV
keywords camouflagedobjectsegmentationzero-shotvision-languagemodelschain-of-thoughtpromptingcross-modalfusionpromptablemedicalimagedepthestimation
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

ArgusCogito claims that camouflaged-object segmentation can be solved without any training on the task by chaining three reasoning steps inside a vision-language model. The pipeline first builds a global scene prior from RGB, depth, and semantic descriptions; then it scans the image in adaptive regions to localize candidates; finally it iteratively sculpts masks by asking the VLM to label fixed point prompts as positive or negative for SAM2 to refine. On four camouflage/plant benchmarks and three medical-image benchmarks, the authors report that this zero-shot chain beats all task-generic prompt baselines and most weakly supervised and fully supervised RGB-D models. The claim matters because it suggests that semantic reasoning over multiple modalities, rather than more supervision or larger segmentation backbones, is what resolves hard foreground-background ambiguity.

What carries the argument

The engine is an iterative point-prompt sculpting loop nested inside a VLM chain-of-thought. Conjecture fuses RGB, depth, scene semantics, candidate regions, and inferred 3D shape into a single prior G3D. Focus adaptively splits the image into N subregions, proposes and verifies candidate bounding boxes, and, when nothing is found, generates camouflage hypotheses and rescans at finer granularity. Sculpting asks SAM2 to emit an initial mask per box, has the VLM diagnose the mask and classify ten regularly spaced points as positive or negative against G3D, then feeds those points back to SAM2; after three rounds the mask stabilizes. The G3D prior and the positive/negative point classification

What would settle it

Pick a random subset of CAMO and COD10K test images and rerun the Sculpting loop three ways: with the VLM's point labels, with labels read from the ground-truth mask at the same ten coordinates, and with randomly shuffled labels. If the ground-truth-label run is not clearly better than the VLM-label run, the VLM's semantic point verification is not doing the claimed work; if the shuffled run retains most of the gain, the loop is mostly SAM2 prompting. Separately, corrupt or permute the depth channel while keeping the chain active: the paper's reasoning predicts a large drop only when reasoning

Watch

Extended reading notes

Core claim

On its own terms, the paper's central claim is that the bottleneck in camouflaged object segmentation is not the segmentation model but the absence of a reasoning process: a vision-language model that conjectures what is in the scene, focuses attention on candidate regions, and then verifies mask structure point by point can turn a generic prompt into a precise mask. With Qwen2.5-VL-7B-Instruct as the reasoning backbone, Depth Anything-v2 for depth maps, and SAM 2.1 Hiera-Large for mask generation, ArgusCogito reports the best task-generic-prompt results on COD10K, CHAMELEON, CAMO, PlantCAMO, CVC-ColonDB, Kvasir, and ISIC, and it beats weakly supervised point- and scribble-based methods and

Load-bearing premise

The load-bearing premise is that the VLM's classification of ten fixed grid points into positive and negative is reliable enough to steer SAM2's masks correctly on every iteration; if those labels are systematically wrong, the loop will amplify rather than correct the error.

Editorial extensions

If this is right

  • Under the paper's claim, task-generic prompting replaces point/scribble supervision for COS; users need only type a phrase like 'camouflaged animal' and supply an RGB image.
  • Depth becomes useful only when mediated by reasoning: raw RGB+depth fusion without the chain gives marginal and unstable gains in the paper's ablation.
  • The same zero-shot chain transfers across domains—animal camouflage, plant camouflage, polyps, skin lesions—with only a changed prompt.
  • Three sculpting iterations are the empirical optimum in the paper's ablation, and the chain's per-image cost scales with the number of VLM calls per round.
  • The framework removes explicit CLIP/Grounding-DINO modular dependencies by using VLM-based point polarity verification inside the Sculpting loop.

Reading between the lines

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

  • The fixed 10-point grid is a crude prompt prior; uncertainty-weighted or mask-adaptive point selection might reach the same mask quality in fewer than three iterations, a testable modification the paper does not explore.
  • Because depth is produced by a monocular estimator, it carries no information fully independent of RGB; an experiment that corrupts or permutes the depth channel would show whether the reported cross-modal gain is true geometric synergy or just extra text grounding.
  • The paper's own ablation says the chain does not transfer to other VLMs; that makes the result a co-design claim about Qwen2.5-VL's grounding abilities as much as about the chain-of-thought, so the practical ceiling depends on backbone grounding quality.
  • If the approach scales, the natural next step is video: the inferred 3D structure M is already a temporal bridge, and the same conjecture-focus-sculpting loop could track moving camouflaged targets.
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 ArgusCogito, a zero-shot chain-of-thought framework for camouflaged object segmentation that combines Qwen2.5-VL, Depth Anything v2, and SAM2.1. It consists of three stages: Conjecture (cross-modal scene/object/semantic reasoning from RGB and depth), Focus (adaptive region decomposition, hypothesis-guided scanning, and candidate verification), and Sculpting (iterative point-prompt refinement with k=3). The authors report state-of-the-art results on COD10K, CHAMELEON, CAMO, PlantCAMO, and three medical-image benchmarks, with ablations on the CAMO dataset and qualitative comparisons to GenSAM, ProMaC, and INT. The central claim is that a zero-shot, prompt-driven VLM/SAM2 pipeline can outperform all weakly supervised and task-generic baselines and rival fully supervised RGB-D models.

Significance. If the reported numbers hold, this is a significant empirical result: it would show that a zero-shot VLM reasoning chain with SAM2 point prompting can substitute for task-specific training and even approach fully supervised RGB-D performance. The paper’s strengths are its breadth of evaluation across seven datasets, comparison with recent methods (ProMaC, INT, GenSAM), and the inclusion of ablations for depth, point verification, and iteration count. However, the central SOTA claim is currently compromised by methodological issues: hyperparameters are selected on the CAMO test set, and no variance or significance information is reported. The manuscript is therefore not yet ready for publication, but the concerns are addressable with a corrected evaluation protocol.

major comments (4)
  1. [§4.2, Table 4, Appendix A.1] The final configuration (focus strategy `auto` and k=3) is selected using the CAMO test set. Table 4(a)–(b) reports ablation results on CAMO, and Appendix A.1 explicitly states that only the 250-image CAMO test split is used, with no validation split described. The final CAMO row in Table 1 is therefore a test-set-selected configuration, not a zero-shot evaluation. Since the same configuration is used for all other datasets, the headline SOTA claim is compromised. The selection margin is small: in Table 4(a), `auto` differs from `single_up` by M 0.079 vs 0.079 and F_beta 0.774 vs 0.761. Please re-run the selection on a held-out validation split or report results for all compared configurations on all datasets, and include run-level variance.
  2. [§4.1] The text states that “All results are averaged over three runs,” but no table reports standard deviations, error bars, or per-run results. Several headline margins are small—e.g., Table 1 CAMO M 0.079 vs 0.086 and F_beta 0.774 vs 0.734 for ArgusCogito vs INT—and Table 4(a) differences are within 0.01–0.02 on most metrics. Without variance estimates it is impossible to determine whether the improvements are statistically meaningful. Please add standard deviations for each metric and, where appropriate, significance tests.
  3. [Appendix C.3] Prompt selection is an additional test-set selection channel. Appendix C.3 states that the best-performing prompt was identified, but does not describe a held-out split or provide a quantitative comparison table of prompt variants. This means the final numbers may include prompt tuning on the test set. Please provide the complete prompt-ablation table with the validation protocol, or alternatively demonstrate that results are stable across all prompt variants.
  4. [§4.2, Table 4(b)] The text says “increasing k leads to consistent improvements,” but Table 4(b) shows M 0.079 at k=1, 0.080 at k=2, and 0.079 at k=3. Only F_beta and the other metrics improve monotonically. This should be rephrased, and the instability in M should be discussed. The absence of variance reporting makes it difficult to interpret these small differences.
minor comments (4)
  1. [Table 3 vs §4.1] Table 3 reports the metric F_w_beta, while Table 1 and Table 4 report F_beta. The metric definitions in §4.1 should explicitly distinguish weighted and unweighted F-measures, and the table headers should be consistent.
  2. [Table 1] The rows `Qwen2.5+SAM2` and `ArgusCogito` are both marked “Ours.” This is confusing; consider renaming the baseline row “Qwen2.5+SAM2 (w/o ArgusCogito)” or similar.
  3. [Appendix C.1] The paragraph claims that the cognitive chain is “not trivially transferable to other vision-language models” and that LLaVA-1.5 is “fundamentally incompatible,” but no quantitative evidence is provided for LLaVA-1.5 or any other VLM. Either add a small comparative experiment or soften the claim.
  4. [Throughout] Several venue names in the tables are misspelled or inconsistent (e.g., “NeurlPS,” “NeruIPS,” “Arxiv,” “AAAI”). Please standardize them. Also, Figures 3 and 4 would benefit from higher-resolution crops, as the qualitative differences are hard to discern.

Circularity Check

1 steps flagged · score 4.0 of 10

CAMO test-set selection of k=3 and the focus strategy makes the CAMO SOTA row a fitted result, not a zero-shot prediction; no other circular steps found.

  1. fitted input called prediction [Section 3.3 'Sculpting' (iteration count k), Section 4.2 Table 4(a)/4(b), Appendix A.1]
    "Based on our ablation study (see Table 4(b)), the number of iterations is empirically set to 3. ... Results in Table 4(b) indicate that increasing k leads to consistent improvements, with optimal performance observed at k = 3. ... the auto configuration achieves the best overall results. ... In the paper, we used the 250 images from its test set."

    Appendix A.1 describes using only the 250-image CAMO test split and no validation split. The ablations in Table 4(a) and 4(b) therefore select both the Focus decomposition mode ('auto') and the Sculpting iteration count (k=3) by optimizing on the CAMO test set. The final ArgusCogito row in Table 1 for CAMO then reports the result of exactly this test-selected configuration. Thus the CAMO SOTA number is not a zero-shot prediction of an untuned pipeline: the optimal k and focus strategy were fitted to the same benchmark that is later announced as the prediction. The reported improvement over k=1/k=2 is the selection criterion, not independent evidence.

full rationale

The equations (1)-(13) define a modular inference pipeline using frozen Qwen2.5-VL, Depth Anything, and SAM2; no model parameter is fitted to the target masks, so the derivation chain itself is not circular. No load-bearing self-citation is present: references to ProMaC, INT, GenSAM, etc. are external works, not the authors' own. The single significant circular step is the evaluation protocol: the CAMO test set is used for hyperparameter selection (focus strategy and k) and then reported as the zero-shot SOTA result. This is a fitted-input-called-prediction issue for the CAMO row and for the headline claim insofar as CAMO is one of the four COS benchmarks. The other benchmarks are less affected because the chosen hyperparameters are fixed and applied to them, although the choice involved peeking at a related COS test set. I therefore assign 4 rather than 6+: the central architecture retains independent content, and the non-CAMO results are not literally constructed from their own test data, but the CAMO evidence for the zero-shot SOTA claim is partially circular. Prompt-sensitivity selection in Appendix C.3 without a stated held-out split is an additional evaluation risk but is not tied to a specific final table, so it is not counted as a separate circular step.

Assumptions & free parameters 5 free parameters · 4 assumptions · 2 invented entities

The framework rests on several pretrained components (Qwen2.5-VL, SAM2.1, Depth Anything-v2, adapted SAM4MLLM) whose reliability is assumed. In addition, k=3, N=3/6, the 10-point grid, the focus strategy, and prompt wording are hand-selected, and k and the focus strategy were tuned on the CAMO test set.

free parameters (5)
  • Number of Sculpting iterations k = 3
    Selected on CAMO test set in Table 4(b) and used for all datasets.
  • Region decomposition N = 3 for Focus; 6 for hypothesis scanning
    Chosen as 'optimizing granularity and semantic alignment' with no ablation; N=6 is added for hypothesis-driven search.
  • Focus strategy = auto
    Ablated on CAMO against single_left, single_up, double, five; 'auto' was selected using test set performance.
  • Point grid size = 10
    A pre-defined set of 10 regularly arranged points in the mask bounding box; no sensitivity study.
  • Prompt templates = hand-authored for Pscene, Pobject, Pinf er, Pdiv, Pf oc, Phyp, Pscan, Pver, Peval, Pgen
    Prompt wording is a tunable design choice; only generic task prompts are shown, not the actual templates.
assumptions (4)
  • domain assumption Qwen2.5-VL-7B reliably performs visual grounding and semantic point classification
    Used in Eqs. 1-9 and 11-12 for all localization, verification, and point classification; no calibration or failure analysis is given.
  • domain assumption Depth maps from Depth Anything-v2 encode usable geometry for camouflaged targets
    Depth maps are passed to the VLM and to SAM2 in Eqs. 1, 2, 5, 6, 8, 9, 10, 12, and 13; Appendix C.2 shows depth alone is unstable, so the synergy must be doing the work.
  • domain assumption SAM2.1 accepts depth as an input
    Eqs. 10 and 13 call SAM2 with IDepth. The public SAM2 interface accepts RGB images and prompts, not a depth channel. Either the notation is wrong or a nonstandard input was used.
  • domain assumption Benchmark annotations and metric implementations are correct
    All quantitative conclusions rest on the ground-truth masks and standard implementations of M, F-beta, E-phi, and S-alpha in COD10K, CHAMELEON, CAMO, PlantCAMO, and the MIS datasets.
invented entities (2)
  • G3D unified feature set (S, R, M, I)
    purpose: Internal prior passed between Conjecture, Focus, and Sculpting
    G3D is an internal textual/multimodal representation with no falsifiable handle outside the paper.
  • Inferred semantic 3D structure M
    purpose: Bridges 2D observation to predicted object morphology in other views
    M is generated by the VLM and is not validated against any 3D ground truth or geometric consistency check.

how reviews work

0 comments
Cite this review

Pith. "Pith review of ArgusCogito: Chain-of-Thought for Cross-Modal Synergy and Omnidirectional Reasoning in Camouflaged Object Segmentation." pith.science (2026). https://pith.science/paper/CZE64HDE

@misc{pith2026250818050,
  author       = {Pith},
  title        = {Pith review of: ArgusCogito: Chain-of-Thought for Cross-Modal Synergy and Omnidirectional Reasoning in Camouflaged Object Segmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CZE64HDE}},
  note         = {Machine review of arXiv:2508.18050}
}
read the original abstract

Camouflaged Object Segmentation (COS) poses a significant challenge due to the intrinsic high similarity between targets and backgrounds, demanding models capable of profound holistic understanding beyond superficial cues. Prevailing methods, often limited by shallow feature representation, inadequate reasoning mechanisms, and weak cross-modal integration, struggle to achieve this depth of cognition, resulting in prevalent issues like incomplete target separation and imprecise segmentation. Inspired by the perceptual strategy of the Hundred-eyed Giant-emphasizing holistic observation, omnidirectional focus, and intensive scrutiny-we introduce ArgusCogito, a novel zero-shot, chain-of-thought framework underpinned by cross-modal synergy and omnidirectional reasoning within Vision-Language Models (VLMs). ArgusCogito orchestrates three cognitively-inspired stages: (1) Conjecture: Constructs a strong cognitive prior through global reasoning with cross-modal fusion (RGB, depth, semantic maps), enabling holistic scene understanding and enhanced target-background disambiguation. (2) Focus: Performs omnidirectional, attention-driven scanning and focused reasoning, guided by semantic priors from Conjecture, enabling precise target localization and region-of-interest refinement. (3) Sculpting: Progressively sculpts high-fidelity segmentation masks by integrating cross-modal information and iteratively generating dense positive/negative point prompts within focused regions, emulating Argus' intensive scrutiny. Extensive evaluations on four challenging COS benchmarks and three Medical Image Segmentation (MIS) benchmarks demonstrate that ArgusCogito achieves state-of-the-art (SOTA) performance, validating the framework's exceptional efficacy, superior generalization capability, and robustness.

Figures

Figures reproduced from arXiv: 2508.18050 by the authors.

Figure 1
Figure 1. Illustration of the human perceptual process in camouflaged object understanding, which [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. The architecture of ArgusCogito, a zero-shot segmentation framework inspired by the perceptual strategy of the Hundred-eyed Giant. It comprises three cognition-driven stages: (1) Conjecture builds a global prior via cross-modal fusion of RGB, depth, and semantics; (2) Focus performs attention-guided, region-level reasoning for precise target localization; (3) Sculpting refines masks iteratively using point-based pro… view at source ↗
Figure 3
Figure 3. Qualitative visualization of camouflaged object segmentation (COS) across six challenging [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figures from the paper (8 more)
Figure 4
Figure 4. Figure 4: Visualization of medical image segmentation (MIS) and plant camouflage detection (PCD), [PITH_FULL_IMAGE:figures/full_fig_p010_4.png]
Figure 5
Figure 5. Figure 5: Effect of removing point verification during the [PITH_FULL_IMAGE:figures/full_fig_p019_5.png]
Figure 6
Figure 6. Figure 6: Segmentation quality over three refinement rounds ( [PITH_FULL_IMAGE:figures/full_fig_p019_6.png]
Figure 7
Figure 7. Figure 7: Qualitative results on the CAMO dataset. [PITH_FULL_IMAGE:figures/full_fig_p020_7.png]
Figure 8
Figure 8. Figure 8: Qualitative results on the CHAMELEON dataset. [PITH_FULL_IMAGE:figures/full_fig_p020_8.png]
Figure 9
Figure 9. Figure 9: Qualitative results on the COD10K dataset. [PITH_FULL_IMAGE:figures/full_fig_p021_9.png]
Figure 10
Figure 10. Figure 10: Qualitative results on the PlantCAMO dataset. [PITH_FULL_IMAGE:figures/full_fig_p021_10.png]
Figure 11
Figure 11. Figure 11: Qualitative results on the MIS task dataset. [PITH_FULL_IMAGE:figures/full_fig_p022_11.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

62 extracted references · 54 canonical work pages

  1. [1]

    Source-free depth for object pop-out

    Zongwei Wu, Danda Pani Paudel, Deng-Ping Fan, Jingjing Wang, Shuo Wang, Cédric De- monceaux, Radu Timofte, and Luc Van Gool. Source-free depth for object pop-out. Jan 2022

  2. [2]

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

    Chunming He, Kai Li, Yachao Zhang, Guoxia Xu, Longxiang Tang, Yulun Zhang, Zhenhua Guo, and Xiu Li. Weakly-supervised concealed object segmentation with sam-based pseudo labeling and multi-scale feature grouping. Advances in Neural Information Processing Systems, 36:30726–30737, 2023

  3. [3]

    Weakly-supervised camouflaged object detection with scribble annotations

    Ruozhen He, Qihua Dong, Jiaying Lin, and Rynson WH Lau. Weakly-supervised camouflaged object detection with scribble annotations. In Proceedings of the AAAI conference on artificial intelligence, volume 37, pages 781–789, 2023

  4. [4]

    Segment anything

    Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4015–4026, 2023

  5. [5]

    Learning transferable visual 10 models from natural language supervision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual 10 models from natural language supervision. In International conference on machine learning, pages 8748–8763. PmLR, 2021

  6. [6]

    Grounding dino: Marrying dino with grounded pre-training for open-set object detection

    Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Qing Jiang, Chunyuan Li, Jianwei Yang, Hang Su, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. In European conference on computer vision, pages 38–55. Springer, 2024

  7. [7]

    Leveraging hallucinations to reduce manual prompt dependency in promptable segmentation

    Jian Hu, Jiayi Lin, Junchi Yan, and Shaogang Gong. Leveraging hallucinations to reduce manual prompt dependency in promptable segmentation. Advances in Neural Information Processing Systems, 37:107171–107197, 2024

  8. [8]

    INT: Instance-Specific Negative Mining for Task-Generic Promptable Segmentation

    Jian Hu, Zixu Cheng, and Shaogang Gong. Int: Instance-specific negative mining for task- generic promptable segmentation. arXiv preprint arXiv:2501.18753, 2025

Show all 62 references
  1. [9]

    Cross-modality salient object detection network with universality and anti-interference

    Hongwei Wen, Kechen Song, Liming Huang, Han Wang, and Yunhui Yan. Cross-modality salient object detection network with universality and anti-interference. Knowledge-Based Systems, 264:110322, Mar 2023

  2. [10]

    M2fnet: Multi-modal fusion network for object detection from visible and thermal infrared images

    Chenchen Jiang, Huazhong Ren, Hong Yang, Hongtao Huo, Pengfei Zhu, Zhaoyuan Yao, Jing Li, Min Sun, and Shihao Yang. M2fnet: Multi-modal fusion network for object detection from visible and thermal infrared images. International Journal of Applied Earth Observation and Geoinfor...

  3. [11]

    Spatial-frequency attention-based optical and scene flow with cross-modal knowledge distillation

    Youjie Zhou, Runyu Jiao, Zhonghan Tao, Xichang Liang, and Yi Wan. Spatial-frequency attention-based optical and scene flow with cross-modal knowledge distillation. The Visual Computer, 41(6):4183–4198, Oct 2024

  4. [12]

    Vision fourier transformer empowered multi-modal imaging system for ethane leakage detection

    Junchi Bin, Shane Rogers, and Zheng Liu. Vision fourier transformer empowered multi-modal imaging system for ethane leakage detection. Information Fusion, 106:102266, Jun 2024

  5. [13]

    Depth- aided camouflaged object detection

    Qingwei Wang, Jinyu Yang, Xiaosheng Yu, Fangyi Wang, Peng Chen, and Feng Zheng. Depth- aided camouflaged object detection. Proceedings of the 31st ACM International Conference on Multimedia, pages 3297–3306, Oct 2023

  6. [14]

    Multi- stream information complementarity network for rgb-d camouflaged object detection

    Chenghao Ying, Zhiping Zhou, Kewei Li, Zhaozhong Zhang, and Qingshuang Yang. Multi- stream information complementarity network for rgb-d camouflaged object detection. The Journal of Supercomputing, 81(5), Apr 2025

  7. [15]

    Depth alignment interaction network for camouflaged object detection

    Hongbo Bi, Yuyu Tong, Jiayuan Zhang, Cong Zhang, Jinghui Tong, and Wei Jin. Depth alignment interaction network for camouflaged object detection. Multimedia Systems, 30(1), Jan 2024

  8. [16]

    Rnet: Multi- modal and multi-scale refined network for rgb-d salient object detection

    Xian Fang, Mingfeng Jiang, Jinchao Zhu, Xiuli Shao, and Hongpeng Wang. Rnet: Multi- modal and multi-scale refined network for rgb-d salient object detection. Pattern Recognition, 135:109139, Mar 2023

  9. [17]

    Cross- modal hierarchical interaction network for rgb-d salient object detection

    Hongbo Bi, Ranwan Wu, Ziqi Liu, Huihui Zhu, Cong Zhang, and Tian-Zhu Xiang. Cross- modal hierarchical interaction network for rgb-d salient object detection. Pattern Recognition, 136:109194, Apr 2023

  10. [18]

    Open-vocabulary camouflaged object segmentation

    Youwei Pang, Xiaoqi Zhao, Jiaming Zuo, Lihe Zhang, and Huchuan Lu. Open-vocabulary camouflaged object segmentation. Jan 2023

  11. [19]

    Exploring deeper! segment anything model with depth perception for camouflaged object detection

    Zhenni Yu, Xiaoqin Zhang, Li Zhao, Yi Bin, and Guobao Xiao. Exploring deeper! segment anything model with depth perception for camouflaged object detection. In Proceedings of the 32nd ACM international conference on multimedia, pages 4322–4330, 2024

  12. [20]

    Weakly-supervised salient object detection via scribble annotations

    Jing Zhang, Xin Yu, Aixuan Li, Peipei Song, Bowen Liu, and Yuchao Dai. Weakly-supervised salient object detection via scribble annotations. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12546–12555, 2020

  13. [21]

    Gernand, Jeffery A

    Yimu Pan, Sitao Zhang, Alison D. Gernand, Jeffery A. Goldstein, and James Z. Wang. Ai-sam: Automatic and interactive segment anything model. Jan 2023. 11

  14. [22]

    Self-prompt sam: Medical image segmentation via automatic prompt sam adaptation

    Bin Xie, Hao Tang, Dawen Cai, Yan Yan, and Gady Agam. Self-prompt sam: Medical image segmentation via automatic prompt sam adaptation. Jan 2025

  15. [23]

    Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models

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

  16. [24]

    Visual instruction tuning.Advances in neural information processing systems, 36:34892–34916, 2023

    Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning.Advances in neural information processing systems, 36:34892–34916, 2023

  17. [25]

    Chain of visual perception: Harnessing multimodal large language models for zero-shot camouflaged object detection

    Lv Tang, Peng-Tao Jiang, Zhihao Shen, Hao Zhang, Jinwei Chen, and Bo Li. Chain of visual perception: Harnessing multimodal large language models for zero-shot camouflaged object detection. Jul 2024

  18. [26]

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

    Jian Hu, Jiayi Lin, Shaogang Gong, and Weitong Cai. Relax image-specific prompt requirement in sam: A single generic prompt for segmenting camouflaged objects. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 12511–12518, 2024

  19. [27]

    Sam 2: Segment anything in images and videos

    Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024

  20. [28]

    How to evaluate foreground maps? In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 248–255, 2014

    Ran Margolin, Lihi Zelnik-Manor, and Ayellet Tal. How to evaluate foreground maps? In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 248–255, 2014

  21. [29]

    Cognitive vision inspired object segmentation metric and loss function

    Deng-Ping Fan, Ge-Peng Ji, Xuebin Qin, and Ming-Ming Cheng. Cognitive vision inspired object segmentation metric and loss function. Scientia Sinica Informationis, 6(6):5, 2021

  22. [30]

    Structure-measure: A new way to evaluate foreground maps

    Deng-Ping Fan, Ming-Ming Cheng, Yun Liu, Tao Li, and Ali Borji. Structure-measure: A new way to evaluate foreground maps. In Proceedings of the IEEE international conference on computer vision, pages 4548–4557, 2017

  23. [31]

    Concealed object detection

    Deng-Ping Fan, Ge-Peng Ji, Ming-Ming Cheng, and Ling Shao. Concealed object detection. IEEE transactions on pattern analysis and machine intelligence, 44(10):6024–6042, 2021

  24. [32]

    Animal camouflage analysis: Chameleon database

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

  25. [33]

    Anabranch network for camouflaged object segmentation

    Trung-Nghia Le, Tam V Nguyen, Zhongliang Nie, Minh-Triet Tran, and Akihiro Sugimoto. Anabranch network for camouflaged object segmentation. Computer vision and image under- standing, 184:45–56, 2019

  26. [34]

    Cross- modality discrepant interaction network for rgb-d salient object detection

    Chen Zhang, Runmin Cong, Qinwei Lin, Lin Ma, Feng Li, Yao Zhao, and Sam Kwong. Cross- modality discrepant interaction network for rgb-d salient object detection. In Proceedings of the 29th ACM international conference on multimedia, pages 2094–2102, 2021

  27. [35]

    Rgb-d saliency detection via cascaded mutual information minimization

    Jing Zhang, Deng-Ping Fan, Yuchao Dai, Xin Yu, Yiran Zhong, Nick Barnes, and Ling Shao. Rgb-d saliency detection via cascaded mutual information minimization. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4338–4347, 2021

  28. [36]

    Specificity- preserving rgb-d saliency detection

    Tao Zhou, Huazhu Fu, Geng Chen, Yi Zhou, Deng-Ping Fan, and Ling Shao. Specificity- preserving rgb-d saliency detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4681–4691, 2021

  29. [37]

    Learning discriminative cross- modality features for rgb-d saliency detection

    Fengyun Wang, Jinshan Pan, Shoukun Xu, and Jinhui Tang. Learning discriminative cross- modality features for rgb-d saliency detection. IEEE Transactions on Image Processing , 31:1285–1297, 2022

  30. [38]

    Improving sam for camouflaged object detection via dual stream adapters

    Jiaming Liu, Linghe Kong, and Guihai Chen. Improving sam for camouflaged object detection via dual stream adapters. arXiv preprint arXiv:2503.06042, 2025

  31. [39]

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

    Siyue Yu, Bingfeng Zhang, Jimin Xiao, and Eng Gee Lim. Structure-consistent weakly supervised salient object detection with local saliency coherence. In Proceedings of the AAAI conference on artificial intelligence, volume 35, pages 3234–3242, 2021. 12

  32. [40]

    Generalized decoding for pixel, image, and language

    Xueyan Zou, Zi-Yi Dou, Jianwei Yang, Zhe Gan, Linjie Li, Chunyuan Li, Xiyang Dai, Harkirat Behl, Jianfeng Wang, Lu Yuan, et al. Generalized decoding for pixel, image, and language. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1511...

  33. [41]

    Segment everything everywhere all at once

    Xueyan Zou, Jianwei Yang, Hao Zhang, Feng Li, Linjie Li, Jianfeng Wang, Lijuan Wang, Jianfeng Gao, and Yong Jae Lee. Segment everything everywhere all at once. Advances in neural information processing systems, 36:19769–19782, 2023

  34. [42]

    Automated polyp detection in colonoscopy videos using shape and context information

    Nima Tajbakhsh, Suryakanth R Gurudu, and Jianming Liang. Automated polyp detection in colonoscopy videos using shape and context information. IEEE transactions on medical imaging, 35(2):630–644, 2015

  35. [43]

    Kvasir-seg: A segmented polyp dataset

    Debesh Jha, Pia H Smedsrud, Michael A Riegler, Pål Halvorsen, Thomas De Lange, Dag Johansen, and Håvard D Johansen. Kvasir-seg: A segmented polyp dataset. In International conference on multimedia modeling, pages 451–462. Springer, 2019

  36. [44]

    Skin lesion analysis toward melanoma detection 2018: A challenge hosted by the international skin imaging collaboration (isic)

    Noel Codella, Veronica Rotemberg, Philipp Tschandl, M Emre Celebi, Stephen Dusza, David Gutman, Brian Helba, Aadi Kalloo, Konstantinos Liopyris, Michael Marchetti, et al. Skin lesion analysis toward melanoma detection 2018: A challenge hosted by the international skin imaging ...

  37. [45]

    Plantcamo: Plant camouflage detection

    Jinyu Yang, Qingwei Wang, Feng Zheng, Peng Chen, Aleš Leonardis, and Deng-Ping Fan. Plantcamo: Plant camouflage detection. arXiv preprint arXiv:2410.17598, 2024

  38. [46]

    Boosting camouflaged object detection with dual-task interactive transformer

    Zhengyi Liu, Zhili Zhang, Yacheng Tan, and Wei Wu. Boosting camouflaged object detection with dual-task interactive transformer. In 2022 26th International Conference on Pattern Recognition (ICPR), pages 140–146. IEEE, 2022

  39. [47]

    High-resolution iterative feedback network for camouflaged object detection

    Xiaobin Hu, Shuo Wang, Xuebin Qin, Hang Dai, Wenqi Ren, Donghao Luo, Ying Tai, and Ling Shao. High-resolution iterative feedback network for camouflaged object detection. In Proceedings of the AAAI Conference on Artificial Intelligence , volume 37, pages 881–889, 2023

  40. [48]

    Vscode: General visual salient and camouflaged object detection with 2d prompt learning

    Ziyang Luo, Nian Liu, Wangbo Zhao, Xuguang Yang, Dingwen Zhang, Deng-Ping Fan, Fahad Khan, and Junwei Han. Vscode: General visual salient and camouflaged object detection with 2d prompt learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognit...

  41. [49]

    Camodiffusion: Camouflaged object detection via conditional diffusion models

    Zhongxi Chen, Ke Sun, and Xianming Lin. Camodiffusion: Camouflaged object detection via conditional diffusion models. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 1272–1280, 2024. 13 Supplementary Material Table of Contents • Experimental ...

  42. [50]

    In the paper, we used 2026 images from its test set

    COD10K: The largest and most comprehensive dataset for camouflaged object detection, containing 10,000 images categorized into 10 superclasses and 78 subclasses, 5,066 images contain camouflaged objects (3,040 for training and 2,026 for testing), accompanied by 1,934 non-camou...

  43. [51]

    It features complex natural backgrounds and manually annotated masks, serving to test generalization in unseen, high-complexity scenes

    CHAMELEON: A small-scale dataset with 76 camouflaged images. It features complex natural backgrounds and manually annotated masks, serving to test generalization in unseen, high-complexity scenes

  44. [52]

    In the paper, we used the 250 images from its test set

    CAMO: Specifically designed for camouflaged object segmentation, comprising 2500 images (1250 camouflaged images from the internet + 1250 from MS-COCO). In the paper, we used the 250 images from its test set. Unlike other datasets, besides natural scenes, the test set of CAMO ...

  45. [53]

    CVC-ColonDB: A classic benchmark for colonic polyp detection, consisting of 300 still frames extracted from 15 colonoscopy videos, each containing a distinct polyp.In this paper, we utilize 380 testing samples

  46. [54]

    Kvasir-SEG: A high-quality dataset focused on polyp segmentation, encompassing diverse polyp shapes, sizes, textures, and realistic imaging artifacts, which present challenges for segmentation tasks.In this paper, we utilized the 100 images from its test set. 14

  47. [55]

    It serves as a standard for dermoscopic segmentation.In this paper, we utilized the 1000 images from its test set

    ISIC 2018: Part of an international benchmark for skin lesion analysis, featuring rigorous protocols, cross-institutional test partitions, and metrics accounting for inter-observer vari- ability. It serves as a standard for dermoscopic segmentation.In this paper, we utilized t...

  48. [56]

    It uses high - resolution, copyright - free images sourced from platforms like Flickr

    PlantCamo: PlantCamo is a novel benchmark designed for plant camouflage detection (PCD). It uses high - resolution, copyright - free images sourced from platforms like Flickr. It features a diverse range of plant species, and its annotations cover both categories and instances...

  49. [57]

    This foundation enables the system to interpret and fuse visual-language information effectively

    Multimodal Reasoning Backbone Qwen2.5-VL-7B-Instruct acts as the multimodal reasoning backbone, facilitating cross- modal semantic understanding across all modules of ArgusCogito. This foundation enables the system to interpret and fuse visual-language information effectively

  50. [58]

    During the Sculpting stage, SAM4MLLM is introduced to perform point-level semantic verification

    Segmentation and Sculpting Module SAM 2.1 Hieral Large is employed for initial segmentation tasks. During the Sculpting stage, SAM4MLLM is introduced to perform point-level semantic verification. This refinement step checks if sampled points lie within target regions, enhancin...

  51. [59]

    These maps enrich spatial perception by providing dense geometric context, complementing RGB information and improving reasoning in cluttered or texture-weak scenes

    Depth Map Pre-generation For each dataset, Depth Anything-v2 is pre-executed to generate depth maps. These maps enrich spatial perception by providing dense geometric context, complementing RGB information and improving reasoning in cluttered or texture-weak scenes

  52. [60]

    (b) Sequential Single-Image Processing : The full workflow (conjecture → focus → sculpting) completes for one image before starting the next

    Chain-of-Thought Execution Modes The conjecture-focus-sculpting pipeline supports two execution strategies for flexibility: (a) Batch Step-by-Step Processing: Modules (conjecture, focus, sculpting) run in batches, processing all images through one module before moving to the n...

  53. [61]

    Eight GPUs are grouped into four pairs (2 GPUs per pair) running in parallel

    GPU Configuration All experiments utilize NVIDIA A100 40G GPUs . Eight GPUs are grouped into four pairs (2 GPUs per pair) running in parallel. This setup accelerates large-scale inference (e.g., full-dataset segmentation) and ablation studies

  54. [62]

    a camouflaged animal hidden in the scene

    Memory Management for High-Resolution Inputs To prevent out-of-memory errors during inference, images with excessively long edges are resized dynamically. The longest side is restricted to 1500 pixels, balancing resolution fidelity and computational feasibility. B Additional Q...

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.