Research map / Detection and segmentation / Open-set and few-shot detection
Shift the Lens: Environment-Aware Unsupervised Camouflaged Object Detection
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 EASE, a training-free unsupervised camouflaged object detection method that builds an environment prototype library (DiffPro: LLaVA-1.5 categories, Stable Diffusion images, DINOv2 features) and segments camouflaged objects by inverting retrieved environment similarity, using KDE-AT, G2L and Self-Retrieval.
Problem
Camouflaged object detection without annotations or training: existing unsupervised methods rely on feature similarity between camouflaged objects and their background, which is highly similar and therefore hard to separate.
Approach
Instead of finding camouflaged objects within the environment, detect the salient environment by retrieving against a library of environment prototypes generated by foundation models (LMM, diffusion, self-supervised ViT), then invert the retrieved environment mask; retrieval uses KDE-based adaptive thresholds, a global-to-local pipeline, and self-retrieval with foreground/background prototypes.
Key results
- On COD10K, EASE (DINOv2-ViT-L/14) reaches S-alpha 0.800 and E-phi 0.884 in unsupervised segmentation, with the paper reporting about a 10% lead over prior unsupervised SOTA on all metrics.
- With SAM-ViT-H prompting, EASE gets 0.856 S-alpha on COD10K, outperforming GenSAM (0.773) and WS-SAM-S (0.803) and comparable to fully-supervised methods such as HitNet (0.871).
- On marine animal segmentation (MAS3K), EASE reaches S-alpha 0.780 versus 0.714 for the best prior unsupervised method TokenCut.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use EASE with DINOv2-ViT-L/14 instead of the strongest listed unsupervised baseline | Unsupervised COD · CHAMELEON | Sα: 0.707 -> 0.819 | Helped |
| Use EASE with DINOv2-ViT-L/14 instead of the strongest listed unsupervised baseline | Unsupervised COD · CAMO | Sα: 0.633 -> 0.749 | Helped |
| Use EASE with DINOv2-ViT-L/14 instead of the strongest listed unsupervised baseline | Unsupervised COD · COD10K | Sα: 0.690 -> 0.773 | Helped |
| Use EASE with DINOv2-ViT-L/14 instead of the strongest listed unsupervised baseline | Unsupervised COD · NC4K | Sα: 0.739 -> 0.800 | Helped |
| Add KDE-AT adaptive thresholding to global retrieval (ablation) | EASE · CHAMELEON | Sα: 0.793 -> 0.806 | Helped |
| Add KDE-AT adaptive thresholding to global retrieval (ablation) | EASE · CAMO | Sα: 0.720 -> 0.729 | Helped |
| Add KDE-AT adaptive thresholding to global retrieval (ablation) | EASE · COD10K | Sα: 0.737 -> 0.749 | Helped |
| Add KDE-AT adaptive thresholding to global retrieval (ablation) | EASE · NC4K | Sα: 0.749 -> 0.763 | Helped |
| Add G2L retrieval after KDE-AT (ablation) | EASE · COD10K | Sα: 0.749 -> 0.763 | Helped |
| Add SR after KDE-AT and G2L (ablation) | EASE · NC4K | Sα: 0.788 -> 0.800 | Helped |
| Replace DiffPro with images crawled from the Web (ablation) | EASE · CAMO | Sα: 0.747 -> 0.749 | No clear effect |
| Replace DiffPro with foreground-masked COD training-set prototypes (ablation) | EASE · COD10K | Sα: 0.759 -> 0.773 | Hurt |
Surprising findings
- Using the DINOv2-ViT-S/14 backbone-based variant, EASE still beats all baselines, and ground-truth training-set-derived prototypes perform worse than generated DiffPro prototypes on several benchmarks.
Limitations
- Performance depends on the quality and coverage of generated environment prototypes and LMM-derived categories.
- Gap to fully-supervised methods remains on CHAMELEON and CAMO for prompt-based variants.
Experimental setup
Datasets: COD10K CAMO LLaVA-1.5 environment categories Stable Diffusion generated environment images
Benchmarks: CHAMELEON CAMO COD10K NC4K MAS3K RMAS500
Metrics: S-measure E-measure weighted F-measure MAE
Baselines: FreeSOLO DiffCut LOST TokenCut Spectral MaskCut FOUND VoteCut MMCPF GenSAM WS-SAM ProMaC
Builds on: LOST TokenCut FOUND DINOv2 SAM
Closest papers by meaning
- Camouflaged Object Detection (CVPR 2020 · similarity 0.92)
- Camouflage Anything: Learning to Hide using Controlled Out-painting and Representation Engineering (CVPR 2025 · similarity 0.88)
- SdalsNet: Self-Distilled Attention Localization and Shift Network for Unsupervised Camouflaged Object Detection (AAAI 2025 · similarity 0.88)
- CamoDiffusion: Camouflaged Object Detection via Conditional Diffusion Models (AAAI 2024 · similarity 0.87)
- UCOD-DPL: Unsupervised Camouflaged Object Detection via Dynamic Pseudo-label Learning (CVPR 2025 · similarity 0.86)
- Endow SAM with Keen Eyes: Temporal-spatial Prompt Learning for Video Camouflaged Object Detection (CVPR 2024 · similarity 0.86)
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.