Research map / Label-efficient and robust learning / OOD and anomaly detection
UniVAD: A Training-free Unified Model for Few-shot Visual Anomaly 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 UniVAD, a training-free unified few-shot visual anomaly detector that uses clustering-refined foundation-model segmentation plus component-aware patch matching and graph-enhanced component modeling to detect industrial, logical, and medical anomalies with a few normal references.
Problem
Visual anomaly detection methods are tailored to single domains (industrial, logical, medical) and usually need many normal samples per category, so they generalize poorly and hinder unified evaluation.
Approach
Combine foundation-model segmentation (RAM, Grounded SAM) with K-means clustering of encoder features to get consistent component masks, then perform patch matching restricted to components and graph-based component matching plus geometric features, with text-based normal/abnormal scoring, all without training on the target domain.
Key results
- Under 1-normal-shot, UniVAD reaches 97.8 image-level AUC on MVTec-AD and 71.0 on MVTec LOCO, versus 84.0 and 62.0 for PatchCore.
- Averaged over domains, UniVAD improves image-level AUC by about 6.2% and pixel-level AUC by 1.7% over the best domain-specific method per dataset.
- Under 4-abnormal-shot medical setting, UniVAD reports 94.1 image-level AUC on BrainMRI and 87.5 on LiverCT, above MVFA (92.4 and 81.2).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use UniVAD instead of the strongest listed baseline | 1-normal-shot visual anomaly detection · MVTec-AD, VisA, MVTec LOCO, BrainMRI, LiverCT, RESC | image-level AUC: MVTec-AD: 94.1 -> 97.8; VisA: 87.4 -> 93.5; MVTec LOCO: 62.2 -> 71.0; BrainMRI: 73.1 -> 80.2; LiverCT: 60.3 -> 70.0; RES | Helped |
| Combine clustering with visual foundation models in C3 instead of using clustering alone (ablation) | C3 module · MVTec-AD | image-level AUC: 97.3 -> 97.8 | Helped |
| Combine clustering with visual foundation models in C3 instead of using visual foundation models alone (ablation) | C3 module · MVTec-AD | pixel-level AUC: 96.1 -> 96.5 | Helped |
| Combine clustering with visual foundation models in C3 instead of using clustering alone (ablation) | C3 module · VisA | image-level AUC: 92.5 -> 93.5 | Helped |
| Add component-aware and image-text matching to patch matching (ablation) | CAPM · MVTec LOCO | image-level AUC: 62.0 -> 64.1 | Helped |
| Add component-aware and image-text matching to patch matching (ablation) | CAPM · BrainMRI | pixel-level AUC: 96.4 -> 96.8 | Helped |
| Add component geometric features (ablation) | GECM · MVTec LOCO | Image-AUC: 64.1 -> 66.6 | Helped |
| Add component geometric features (ablation) | GECM · MVTec LOCO | Pixel-AUC: 70.2 -> 73.2 | Helped |
| Add component deep features to geometric features (ablation) | GECM · MVTec LOCO | Image-AUC: 66.6 -> 69.4 | Helped |
| Add component feature aggregation (CFA) (ablation) | GECM · MVTec LOCO | Image-AUC: 69.4 -> 71.0 | Helped |
| Add component feature aggregation (CFA) (ablation) | GECM · MVTec LOCO | Pixel-AUC: 74.8 -> 75.1 | Helped |
| Use CLIP-ViT, DINO, or DINOv2 image features for clustering (ablation) | C3 clustering · MVTec-AD, VisA, MVTec LOCO, RESC | image-level AUC | No clear effect |
Surprising findings
- Standard PatchCore drops from 84.1% image AUC on MVTec-AD to 62.0% on MVTec LOCO, showing domain-specific methods do not transfer.
Limitations
- Relies on multiple large visual foundation models (RAM, Grounding DINO, SAM), causing inference latency that limits real-time use.
- Hyperparameters such as thresholds and fusion weights are set manually rather than learned.
Experimental setup
Datasets: MVTec-AD VisA MVTec LOCO BrainMRI LiverCT RetinalOCT ChestXray HIS OCT17
Benchmarks: MVTec-AD VisA MVTec LOCO BrainMRI LiverCT RESC HIS ChestXray OCT17
Metrics: image-level AUC pixel-level AUC
Baselines: PatchCore AnomalyGPT WinCLIP ComAD UniAD MedCLIP DRA BGAD MVFA
Builds on: PatchCore WinCLIP ComAD SAM Grounded SAM
Closest papers by meaning
- One-to-Normal: Anomaly Personalization for Few-shot Anomaly Detection (NeurIPS 2024 · similarity 0.88)
- Advancing Video Anomaly Detection: A Concise Review and a New Dataset (NeurIPS 2024 · similarity 0.88)
- Towards Zero-Shot Anomaly Detection and Reasoning with Multimodal Large Language Models (CVPR 2025 · similarity 0.88)
- Toward Generalist Anomaly Detection via In-context Residual Learning with Few-shot Sample Prompts (CVPR 2024 · similarity 0.88)
- Open-Vocabulary Video Anomaly Detection (CVPR 2024 · similarity 0.88)
- AnomalyNCD: Towards Novel Anomaly Class Discovery in Industrial Scenarios (CVPR 2025 · similarity 0.87)
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.