Research map / Detection and segmentation / Open-set and few-shot detection
FlexUOD: The Answer to Real-world Unsupervised Image Outlier 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 FlexUOD, a training-free unsupervised image outlier detection framework that estimates the contamination factor via Spearman rank similarity between a low-contamination detector and a high-contamination detector, then selects/blends them accordingly.
Problem
Unsupervised outlier detection (UOD) methods perform unstably across varying contamination factors (outlier ratios) in unlabeled datasets, and existing contamination factor estimators based on threshold learning are disconnected from the outlier score function.
Approach
Builds structurally consistent low-contamination (hypersphere distance with Bray-Curtis) and high-contamination (shell normalization over MAD-filtered outlier candidates) scores, then uses the Spearman rank correlation between them as an outlier-detector-aware estimate of the contamination factor to choose the scoring method and threshold.
Key results
- Averaged over outlier ratios, FlexUOD reaches AUC 0.980 on STL-10 and 0.942 on CIFAR-10 (with ResNet-50 and CLIP averaged), beating the prior SOTA by about 6% on CIFAR-10.
- On threshold F1-score, FlexUOD reports 0.687 on STL-10 versus 0.660 for the best threshold-learning baseline KARCH.
- Plugging the estimated contamination-aware selection into OCSVM, LVAD and ICL improves their average AUC and stability on STL-10 (Fig. 8).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use FlexUOD instead of Multi-T | UOD with ResNet-50 and CLIP average features · STL-10 | AUC: 0.957 -> 0.980 | Helped |
| Use FlexUOD instead of Multi-T | UOD with ResNet-50 and CLIP average features · Internet | AUC: 0.956 -> 0.981 | Helped |
| Use FlexUOD instead of Multi-T | UOD with ResNet-50 and CLIP average features · CIFAR-10 | AUC: 0.888 -> 0.942 | Helped |
| Use FlexUOD instead of Multi-T | UOD with ResNet-50 and CLIP average features · CIFAR-100 | AUC: 0.917 -> 0.947 | Helped |
| Use FlexUOD instead of LVAD | UOD with ResNet-50 and CLIP average features · MIT-Places-Small | AUC: 0.860 -> 0.928 | Helped |
| Add Bray-Curtis distance (ablation) | Baseline Eq. 5 · STL-10 | AUC: 0.950 -> 0.963 (+1.4%) | Helped |
| Add the full FlexUOD framework (ablation) | Baseline Eq. 5 · STL-10 | AUC: 0.950 -> 0.980 (+3.2%) | Helped |
| Add Bray-Curtis distance (ablation) | Baseline Eq. 5 · CIFAR-10 | AUC: 0.870 -> 0.919 (+5.6%) | Helped |
| Add the full FlexUOD framework (ablation) | Baseline Eq. 5 · CIFAR-10 | AUC: 0.870 -> 0.942 (+8.3%) | Helped |
| Add Bray-Curtis distance (ablation) | Baseline Eq. 5 · CIFAR-100 | AUC: 0.891 -> 0.933 (+4.7%) | Helped |
| Add the full FlexUOD framework (ablation) | Baseline Eq. 5 · CIFAR-100 | AUC: 0.891 -> 0.947 (+6.3%) | Helped |
| Add Bray-Curtis distance (ablation) | Baseline Eq. 5 · MNIST | AUC: 0.861 -> 0.877 (+1.9%) | Helped |
Surprising findings
- A simple hypersphere-distance baseline without any feature refinement already performs competitively with deep methods such as LVAD across contamination factors.
- Industrial defect detection (MVTec-AD) is treated as patch-level rather than semantic outlier detection, and the unsupervised method is not the best there for many categories.
Limitations
- May be less effective on anomaly segmentation/localization tasks since it operates on whole-image features.
- Fixed Spearman boundaries (0.1, 0.3) and the choice of omega and MAD parameter k are heuristic, selected by ablation.
Experimental setup
Datasets: STL-10 Internet CIFAR-10 CIFAR-100 MIT-Places-Small MNIST Fashion-MNIST MVTec-AD BMAD (HIS, BrainMRI, LiverCT, RESC) ADBench
Benchmarks: STL-10 Internet CIFAR-10 CIFAR-100 MIT-Places-Small MNIST Fashion-MNIST MVTec-AD BMAD medical-AD ADBench tabular-AD
Metrics: AUC AUPR-In AUPR-Out F1-score
Baselines: OC-SVM Isolation Forest INNE Deep SVDD RSRAE GOAD ICL LVAD Shell-Re Multi-T KARCH REGR
Builds on: Shell theory (Shell-Re) LVAD MAD outlier filtering Bray-Curtis distance
Closest papers by meaning
- ODIM: Outlier Detection via Likelihood of Under-Fitted Generative Models (ICML 2024 · similarity 0.86)
- RODEO: Robust Outlier Detection via Exposing Adaptive Out-of-Distribution Samples (ICML 2024 · similarity 0.85)
- How Does Unlabeled Data Provably Help Out-of-Distribution Detection? (ICLR 2024 · similarity 0.85)
- Robust Conformal Outlier Detection under Contaminated Reference Data (ICML 2025 · similarity 0.84)
- ALTBI: Constructing Improved Outlier Detection Models via Optimization of Inlier-Memorization Effect (AAAI 2025 · similarity 0.84)
- UniVAD: A Training-free Unified Model for Few-shot Visual Anomaly Detection (CVPR 2025 · 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.