MVTec AD
Overview
MVTec AD is the gold-standard benchmark for industrial anomaly detection. It contains 15 categories of objects and textures commonly found in manufacturing, with pixel-accurate defect annotations.
What sets MVTec AD apart from synthetic benchmarks is that the defects are real physical anomalies — not injected noise. This makes it a genuinely hard problem that mirrors actual factory conditions.
Dataset Structure
mvtec_ad/
├── bottle/
│ ├── train/good/ # defect-free images
│ └── test/
│ ├── good/
│ └── broken/ # defective images
│ └── contaminated/
├── cable/
│ └── ...
└── [15 categories total]
Each category has:
- Training set: only defect-free images (normal)
- Test set: mix of good and defective images with per-pixel masks
Categories
The 15 categories span both objects and textures:
| Category | Type | Defect Types |
|---|---|---|
| bottle | object | broken, contaminated |
| cable | object | bent, broken, cut, missing_cable |
| capsule | object | cracked, dirty, scratch |
| carpet | texture | hole, metal_contamination, thread |
| grid | texture | broken, flat |
| hazelnut | object | crack, hole, print |
| leather | texture | cut, fold, glue, poke |
| metal_nut | object | bent, broken, flip, scratch |
| pill | object | color, contamination, cracked, flaky, pitted |
| screw | object | manipulated_front, scratch, thread_side |
| tile | texture | crack, glaze, print |
| toothbrush | object | bent, broken |
| transistor | object | bent_lead, cut_lead, damaged_case |
| wood | texture | hole, liquid, color |
| zipper | object | broken_teeth, fabric_border, rough |
Metrics
MVTec AD is evaluated on two levels:
| Level | Metric | Description |
|---|---|---|
| Image | AUROC | Is the image defective or not? |
| Pixel | AUROC | Where is the defect located? |
Pixel-level evaluation requires per-pixel anomaly masks, making it a much harder task than image-level classification.
Why It Matters
Before MVTec AD, anomaly detection papers used ad-hoc datasets that were hard to reproduce. MVTec AD gave the community:
- Consistent benchmark across 15 diverse categories
- Real defects from actual manufacturing processes
- Pixel-accurate ground truth for localization
- Train/test splits that reflect how you'd actually deploy a system
Access
- Website: MVTec AD Dataset
- Download: Registration required (free for academic use)
- Kaggle: Available via Kaggle datasets
Citation
If you use MVTec AD in published work:
@misc{mvtecad,
title={MVTec AD — A Comprehensive Real-World Dataset for Unsupervised Anomaly Detection},
author={Paul Bergmann and Michael Fauser and David Sattlegger and Carsten Steger},
year={2019},
booktitle={CVPR}
}
Related
- PatchCore Anomaly Detection — my from-scratch implementation evaluated on MVTec AD