Research map / Detection and segmentation / CNN features
Strategic Preys Make Acute Predators: Enhancing Camouflaged Object Detectors by Generating Camouflaged Objects
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes Camouflageator, an adversarial generator-detector training framework that synthesizes harder camouflaged objects, and ICEG, a COD detector with camouflaged feature coherence and edge-guided separated calibration modules, reaching state-of-the-art COD results.
Problem
Camouflaged object detection struggles to produce complete segmentations and sharp boundaries in extreme camouflage cases, and existing detectors generalize poorly to hard camouflaged examples.
Approach
Prey side: an auxiliary generator is adversarially trained against the detector to synthesize harder camouflaged objects (with fidelity and concealment losses). Predator side: ICEG adds a camouflaged consistency loss to enforce internal coherence and an edge-guided separated calibration module that splits foreground/background attentive features and uses edge-based adaptive normalization.
Key results
- ICEG with Res2Net50 backbone outperforms BGNet on COD10K (M 0.028 vs 0.033, F-beta 0.752 vs 0.714) and ICEG with Swin backbone surpasses ICON by about 8.3% in the common setting.
- Camouflageator improves existing detectors: PreyNet by 2.8%, FGANet by 2.2%, FEDER by 2.3%, and ICEG (ResNet50) by 2.5%.
- On COD10K, ablation shows the camouflaged consistency loss improves ICEG by about 3.4%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add Camouflageator training | ICEG with ResNet50 · CHAMELEON | Fβ: 0.858 -> 0.863 | Helped |
| Add Camouflageator training | ICEG with ResNet50 · CAMO | Fβ: 0.789 -> 0.805 | Helped |
| Add Camouflageator training | ICEG with ResNet50 · COD10K | Fβ: 0.747 -> 0.763 | Helped |
| Add Camouflageator training | ICEG with ResNet50 · NC4K | Fβ: 0.814 -> 0.835 | Helped |
| Add CFC (ablation) | ICEG · COD10K | Fβ: 0.685 -> 0.747 | Helped |
| Add IFA (ablation) | ICEG · COD10K | Fβ: 0.728 -> 0.747 | Helped |
| Add CFA (ablation) | ICEG · COD10K | Fβ: 0.731 -> 0.747 | Helped |
| Add feature aggregation (ablation) | ICEG · COD10K | Fβ: 0.720 -> 0.747 | Helped |
| Add camouflaged consistency loss (ablation) | ICEG · COD10K | Fβ: 0.722 -> 0.747 | Helped |
| Use the full camouflaged consistency loss instead of only its first term (ablation) | ICEG · COD10K | Fβ: 0.704 -> 0.747 | Helped |
| Add ESD (ablation) | ICEG · COD10K | Fβ: 0.678 -> 0.747 | Helped |
| Add ESC (ablation) | ICEG · COD10K | Fβ: 0.688 -> 0.747 | Helped |
Limitations
- Camouflageator requires an extra pre-training stage and adversarial training, adding training complexity
- Reported improvements are measured only under the common COD setting and image size 352x352
Experimental setup
Datasets: CAMO COD10K CHAMELEON NC4K ImageNet
Benchmarks: CHAMELEON CAMO COD10K NC4K
Metrics: mean absolute error adaptive F-measure mean E-measure structure measure
Baselines: SegMaR PreyNet FGANet FEDER SINet V2 BGNet ICON ZoomNet
Builds on: PreyNet FEDER SegMaR ZoomNet bi-level optimization
Closest papers by meaning
- Simultaneously Localize, Segment and Rank the Camouflaged Objects (CVPR 2021 · similarity 0.90)
- Camouflaged Object Detection (CVPR 2020 · similarity 0.87)
- CamoDiffusion: Camouflaged Object Detection via Conditional Diffusion Models (AAAI 2024 · similarity 0.85)
- Shift the Lens: Environment-Aware Unsupervised Camouflaged Object Detection (CVPR 2025 · similarity 0.84)
- Camouflage Anything: Learning to Hide using Controlled Out-painting and Representation Engineering (CVPR 2025 · similarity 0.84)
- CNCA: Toward Customizable and Natural Generation of Adversarial Camouflage for Vehicle Detectors (NeurIPS 2024 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.