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

Jianwen Tan1,2,* Huiyao Zhang1,2,* Rui Xiong1,2 Han Zhou1,2 Hongfei Wang2 Ye Li2,†
1 University of Chinese Academy of Sciences 2 Technology and Engineering Center for Space Utilization, Chinese Academy of Sciences
* Equal contribution    † Corresponding author
arXiv 2025
Human perceptual process for camouflaged object understanding and the three-stage ArgusCogito pipeline

Figure 1. Human observation with RGB alone is often insufficient. Depth-assisted perception and the Conjecture-Focus-Sculpting chain progressively reveal the hidden object.

ArgusCogito is a zero-shot segmentation framework inspired by the perceptual strategy of the Hundred-eyed Giant. It reformulates camouflaged object segmentation as a cohesive reasoning process and has the following properties:

Method Overview

ArgusCogito organizes multimodal reasoning into three cognition-driven stages, preserving semantic information from global scene understanding to pixel-level mask correction.

Architecture of ArgusCogito with Conjecture, Focus, and Sculpting
Figure 2. The full ArgusCogito architecture. Conjecture forms a global multimodal prior; Focus locates all plausible instances through adaptive scanning; Sculpting refines masks with semantic point verification.

1 Conjecture

Given an RGB image and a depth map, the VLM constructs scene semantics, candidate object concepts, inferred 3D structure, and camouflage mechanisms. These elements form the global prior G3D.

2 Focus

The scene is adaptively divided along horizontal or vertical directions. Semantic focusing, hypothesis-driven reasoning, and candidate verification recover objects that coarse global reasoning may miss.

3 Sculpting

SAM 2.1 produces an initial mask. The VLM diagnoses semantic and structural errors, verifies dense positive/negative points, and updates the mask for three refinement rounds.

Quantitative Results on Camouflaged Object Segmentation

Under the task-generic prompt setting, ArgusCogito consistently improves over prompt-driven baselines and rivals several supervised RGB-D methods. Lower M is better; higher Fβ, Eϕ, and Sα are better.

Method COD10K CHAMELEON CAMO
M ↓Fβ ↑Eϕ ↑Sα ↑ M ↓Fβ ↑Eϕ ↑Sα ↑ M ↓Fβ ↑Eϕ ↑Sα ↑
Qwen2.5 + SAM2.126.656.792.734.086.625.786.728.127.591.728.681
GroundingSAM.085.670.813.764.122.662.776.744.157.656.753.707
GenSAM.058.695.843.783.073.696.806.774.106.669.798.729
ProMaC.042.716.876.805.044.790.899.833.090.725.846.767
INT.037.722.883.808.039.801.906.842.086.734.853.772
ArgusCogito.026.824.928.843.035.824.918.859.079.774.866.800

Table 1. Representative task-generic prompt baselines from the paper. All results are averaged over three runs.

Comparison on Challenging Camouflage Scenarios

The comparison covers indefinable boundaries, small objects, large objects, occlusion, multiple objects, and complex shapes. ArgusCogito improves both instance completeness and boundary fidelity.

Qualitative comparison against GenSAM and ProMaC across six camouflage challenges
Figure 3. Qualitative COS comparison across six challenging scenarios. The proposed reasoning chain reduces semantic drift, omissions, and fragmented masks.

Cross-Domain Zero-Shot Generalization

The same task-generic reasoning pipeline transfers from natural camouflage to plant camouflage and medical imagery without task-specific fine-tuning.

Qualitative results on plant camouflage detection and medical image segmentation
Figure 4. Plant camouflage detection, polyp segmentation, and skin lesion segmentation compared with GenSAM and ProMaC.
DatasetDomainTest samplesM ↓Fβ ↑Eϕ ↑Sα ↑
CVC-ColonDBPolyp380.106.518.712.685
KvasirPolyp100.086.725.828.794
ISICSkin lesion1,000.121.782.813.777
PlantCAMOPlant camouflage250.112.597a.779.694

a PlantCAMO reports weighted F-measure Fwβ. Dataset split sizes follow the supplementary material.

Ablation Studies

The supplementary experiments separate the contribution of the cognitive chain from the underlying backbone and raw depth cues.

Backbone vs. reasoning chain

SettingM ↓Fβ ↑Eϕ ↑Sα ↑
Qwen2.5 + SAM2.126.656.792.734
ArgusCogito.026.824.928.843

The gains are not explained by the foundation models alone. The structured cognitive chain is required to unlock fine-grained spatial grounding.

Contribution of depth

RGBDepthReasoningM ↓Fβ ↑Sα ↑
✓––.126.656.734
✓✓–.056.628.742
✓✓✓.026.824.843

Depth alone is unstable. It becomes useful when the reasoning chain interprets it as a context-dependent geometric scaffold.

Point verification and iterative sculpting

Effect of semantic point verification during Sculpting
Figure 5. Without point verification, masks exhibit spatial leakage and semantic drift near ambiguous boundaries.
Segmentation quality over one, two, and three refinement rounds
Figure 6. Three rounds progressively improve boundary alignment and mask completeness.

Dynamic Focus strategy on CAMO

StrategyM ↓Fβ ↑Eϕ ↑Sα ↑
single-left.081.758.862.797
single-up.079.761.863.797
double.083.752.859.798
five.085.746.862.795
adaptive.079.774.866.800

Number of Sculpting iterations

IterationsM ↓Fβ ↑Eϕ ↑Sα ↑
1.079.759.860.795
2.080.765.862.794
3.079.774.866.800

Prompt sensitivity

Task-aligned prompts that encode structural or biological priors are the most stable. Prompts emphasizing full-body coverage, reduced fragmentation, or domain-specific cues improve reasoning consistency; vague conversational requests often fail to trigger effective spatial abstraction. The analysis suggests that semantic precision matters more than prompt length.

Evaluation Datasets

Seven benchmarks cover natural camouflage, plant camouflage, polyp segmentation, and skin lesion segmentation.

COD10K

10,000 images across 10 superclasses and 78 subclasses; 2,026 test images are evaluated.

CHAMELEON

76 complex natural images for testing generalization in unseen high-ambiguity scenes.

CAMO

2,500 images; the 250-image test split includes natural camouflage and body painting.

CVC-ColonDB

Colonoscopy frames with diverse polyps, evaluated under the PraNet partition protocol.

Kvasir-SEG

High-quality polyp masks spanning varied shapes, textures, sizes, and imaging artifacts.

ISIC 2018

A standard dermoscopic skin-lesion benchmark with cross-institutional variation.

PlantCAMO

High-resolution, instance-annotated plant camouflage images with strong inter-class similarity.

Additional Qualitative Results

The supplementary material provides full comparisons with GenSAM and ProMaC across five representative benchmarks.

CAMO

Global full-resolution reasoning and iterative refinement improve boundary precision and shape completeness in visually homogeneous scenes.

Qualitative comparison on the CAMO dataset
Figure 7. Qualitative results on CAMO.

CHAMELEON

Dynamic spatial focus suppresses false positives caused by subtle texture and luminance differences while preserving complete object structure.

Qualitative comparison on the CHAMELEON dataset
Figure 8. Qualitative results on CHAMELEON.

COD10K

Task-generic prompting and the structured reasoning chain remain robust across rare categories and complex natural backgrounds.

Qualitative comparison on the COD10K dataset
Figure 9. Qualitative results on COD10K.

PlantCAMO

Geometry, semantics, and spatial layout enable domain transfer to fine-grained plant camouflage without specialized adaptation.

Qualitative comparison on the PlantCAMO dataset
Figure 10. Qualitative results on PlantCAMO.

Medical Image Segmentation

Prompt-informed spatial reasoning preserves coherence in polyps and skin lesions despite low contrast, flattened geometry, and unfamiliar clinical imagery.

Qualitative comparison on polyp and skin lesion segmentation
Figure 11. Qualitative results on medical image segmentation.

Abstract

Camouflaged Object Segmentation (COS) is challenging because targets and backgrounds are highly similar in color, texture, and structure. Existing methods often rely on shallow feature representations, limited reasoning, or weak cross-modal integration, leading to missed instances and inaccurate boundaries. Inspired by the perceptual strategy of the Hundred-eyed Giant, we introduce ArgusCogito, a zero-shot chain-of-thought framework for cross-modal synergy and omnidirectional reasoning within Vision-Language Models.

ArgusCogito contains three cognition-inspired stages. Conjecture constructs a global cognitive prior through RGB-depth-semantic reasoning. Focus performs adaptive omnidirectional scanning and candidate verification. Sculpting generates high-fidelity masks through iterative semantic feedback and dense positive/negative point prompts. Evaluations across camouflaged object segmentation, medical image segmentation, and plant camouflage detection demonstrate strong accuracy, robustness, and cross-domain generalization.

Implementation Details

Reasoning backbone

Qwen2.5-VL-7B-Instruct provides multimodal scene understanding, fine-grained visual grounding, and point-wise semantic verification.

Segmentation

SAM 2.1 Hiera Large generates initial masks. A SAM4MLLM-derived binary classifier verifies point prompts during Sculpting.

Depth input

Depth Anything V2 pre-generates monocular depth maps, adding dense geometric context for low-contrast or cluttered scenes.

Inference

The pipeline supports batched stage-wise execution and sequential single-image execution. Images are resized to a maximum long edge of 1,500 pixels.

Hardware

Experiments use NVIDIA A100 40 GB GPUs, organized as four two-GPU pairs for large-scale inference and ablation studies.

Prompts

Task-generic prompts identify camouflaged animals, camouflaged plants, polyps, or skin lesions without spatial priors or task-specific training.

Discussion and Future Directions

More modalities

Infrared and thermal inputs could improve robustness in low-light, occluded, or safety-critical environments.

Adaptive granularity

Dynamic zoom-in and hierarchical decomposition may better handle ultra-small targets and dense-object scenes.

Domain adaptation

Lightweight fine-tuning or a dynamic camouflage-pattern lexicon could inject specialized domain knowledge.

Broader scenarios

Industrial defect detection and temporally consistent video segmentation are promising extensions.

BibTeX

@article{tan2025arguscogito,
  title   = {ArgusCogito: Chain-of-Thought for Cross-Modal Synergy and
             Omnidirectional Reasoning in Camouflaged Object Segmentation},
  author  = {Tan, Jianwen and Zhang, Huiyao and Xiong, Rui and Zhou, Han
             and Wang, Hongfei and Li, Ye},
  journal = {arXiv preprint arXiv:2508.18050},
  year    = {2025}
}