Research map / Detection and segmentation / Open-set and few-shot detection
CamoDiffusion: Camouflaged Object Detection via Conditional Diffusion Models
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.
Treats camouflaged object detection as conditional mask generation with a diffusion model, adding a PVTv2-based conditional network, structure corruption, SNR-shifted variance schedule, and consensus time ensemble sampling.
Problem
Camouflaged object detection struggles with nuanced object boundaries and overconfident incorrect point-estimate predictions from standard segmentation models.
Approach
Reformulates COD as conditional diffusion mask generation: an Adaptive Transformer Conditional Network (PVT with zero-overlapping embedding and time-token concatenation) supplies image conditions per denoising step, trained with structure corruption and an SNR-shifted variance schedule, and sampled with a consensus time ensemble over denoising steps.
Key results
- On COD10K, CamoDiffusion reaches MAE 0.019 (CamoDiffusion-E) and 0.020 (CamoDiffusion), reducing MAE by 20.9% and raising weighted F by 7.7% over the second-best FSPNet.
- On CAMO, CamoDiffusion achieves S-alpha 0.878 and MAE 0.042 versus HitNet's 0.849 and 0.055 (as listed in Table 1 comparisons).
- Ablation on COD10K shows ZOE gives the largest gain; CTE reduces the share of overconfident incorrect pixels among incorrect pixels from 91% to 76%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use CamoDiffusion instead of the strongest reported baseline | COD10K · COD10K | M: 0.023 -> 0.020 | Helped |
| Use CamoDiffusion instead of the strongest reported baseline | CAMO · CAMO | Sα: 0.856 -> 0.878 | Helped |
| Use CamoDiffusion instead of the strongest reported baseline | NC4K · NC4K | F w β: 0.834 -> 0.859 | Helped |
| Add ATCN and TTC (ablation) | CamoDiffusion · COD10K | Sα: 0.867 -> 0.881 | Helped |
| Add Zero Overlapping Embedding (ablation) | CamoDiffusion · COD10K | Sα: 0.872 -> 0.881 | Helped |
| Add Time Token Concatenation (ablation) | CamoDiffusion · COD10K | Sα: 0.877 -> 0.881 | Helped |
| Add the SNR-based variance schedule and Structure Corruption (ablation) | CamoDiffusion · COD10K | Sα: 0.821 -> 0.881 | Helped |
| Add the SNR-based variance schedule (ablation) | CamoDiffusion · COD10K | Sα: 0.868 -> 0.881 | Helped |
| Add Structure Corruption (ablation) | CamoDiffusion · COD10K | Sα: 0.855 -> 0.881 | Helped |
| Apply CTE (ablation) | CamoDiffusion · COD10K | M: 0.021 -> 0.020 | Helped |
| Ensemble three sampling runs with CTE | CamoDiffusion · COD10K | M: 0.020 -> 0.019 | Helped |
Surprising findings
- Removing ZOE from ATCN yields the largest drop in some ablations, while dropping the entire ATCN with no ZOE or TTC degrades results substantially (S-alpha 0.867 vs 0.881 on COD10K).
Limitations
- Sampling requires multiple denoising steps, and CamoDiffusion-E triples inference passes.
- Evaluated only on three COD datasets with image-only input.
Experimental setup
Datasets: CAMO COD10K NC4K
Benchmarks: CAMO COD10K NC4K
Metrics: S-measure E-measure weighted F-measure MAE
Baselines: SINet-V2 ZoomNet DTINet HitNet FSPNet DGNet EnsemDiff MedSegDiff UGTR PreyNet
Builds on: DDPM PVT Local Emphasis SNR-based variance schedule (Hoogeboom et al. 2023) Bit diffusion-style segmentation
Closest papers by meaning
- Shift the Lens: Environment-Aware Unsupervised Camouflaged Object Detection (CVPR 2025 · similarity 0.87)
- Camouflage Anything: Learning to Hide using Controlled Out-painting and Representation Engineering (CVPR 2025 · similarity 0.87)
- UCOD-DPL: Unsupervised Camouflaged Object Detection via Dynamic Pseudo-label Learning (CVPR 2025 · similarity 0.86)
- ReDiffDet: Rotation-equivariant Diffusion Model for Oriented Object Detection (CVPR 2025 · similarity 0.86)
- DiMSOD: A Diffusion-Based Framework for Multi-Modal Salient Object Detection (AAAI 2025 · similarity 0.85)
- Strategic Preys Make Acute Predators: Enhancing Camouflaged Object Detectors by Generating Camouflaged Objects (ICLR 2024 · similarity 0.85)
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.