Skip to content
intermediateimage

MVTec AD Industrial Defect Detection

Anomaly detection and localization in industrial images (Bergmann et al., CVPR 2019). Detect defects across 15 object and texture categories including metal nuts, transistors, and leather.

About this dataset

Anomaly detection asks a model to flag images that deviate from a set of known-good examples, and to localize where the deviation is. The MVTec Anomaly Detection (MVTec AD) dataset was introduced by Paul Bergmann, Michael Fauser, David Sattlegger, and Carsten Steger at CVPR 2019 to give unsupervised anomaly detection a realistic industrial benchmark, since earlier datasets were small and did not reflect factory inspection conditions.

The images show manufactured objects and material textures photographed under controlled lighting. Training images are all defect-free, and the test images add real manufacturing defects such as scratches, dents, contaminations, and structural changes. Each defect in the test set comes with a pixel-precise ground-truth region marking exactly where the anomaly is.

The dataset contains 5,354 high-resolution color images spread across 15 categories, split into 10 object categories and 5 texture categories. The test images cover over 70 different types of defects, and every anomaly is labeled with a pixel-precise ground-truth mask. Because the training split holds only normal samples, the benchmark is aimed at methods that learn what 'good' looks like without seeing defects during training.

The Potato config below reproduces this task with a radio question for the product category (the 15 MVTec classes), a radio question for normal versus defective, a multiselect for the defect types present, and a free-text field for describing where the defect sits. This turns MVTec AD into an image-level triage and defect-tagging pass that a human can run over inspection photos.

Total images
5,354
Categories
15 (10 objects, 5 textures)
Defect types
Over 70
Ground truth
Pixel-precise anomaly regions
Training data
Defect-free images only
Venue
CVPR 2019
Labels:outdoornatureurbanpeopleanimal+−

Configuration Fileconfig.yaml

This Potato config reproduces the annotation task. Save it as config.yaml and run potato start config.yaml to try it.

yaml
# MVTec AD Industrial Defect Detection Configuration
# Based on Bergmann et al., CVPR 2019

annotation_task_name: "MVTec AD Defect Detection"
task_dir: "."

data_files:
  - "sample-data.json"

item_properties:
  id_key: "id"
  text_key: "image_url"
  context_key: "product_type"

user_config:
  allow_all_users: true

annotation_schemes:
  - annotation_type: "radio"
    name: "product_category"
    description: "Select the product category"
    labels:
      - name: "bottle"
        tooltip: "Glass or plastic bottles"
      - name: "cable"
        tooltip: "Electrical cables"
      - name: "capsule"
        tooltip: "Capsules, pills"
      - name: "carpet"
        tooltip: "Carpet texture"
      - name: "grid"
        tooltip: "Grid patterns"
      - name: "hazelnut"
        tooltip: "Hazelnuts"
      - name: "leather"
        tooltip: "Leather texture"
      - name: "metal_nut"
        tooltip: "Metal nuts"
      - name: "pill"
        tooltip: "Pills"
      - name: "screw"
        tooltip: "Screws"
      - name: "tile"
        tooltip: "Tiles"
      - name: "toothbrush"
        tooltip: "Toothbrushes"
      - name: "transistor"
        tooltip: "Transistors"
      - name: "wood"
        tooltip: "Wood texture"
      - name: "zipper"
        tooltip: "Zippers"

  - annotation_type: "radio"
    name: "anomaly_status"
    description: "Is this image normal or defective?"
    labels:
      - name: "good"
        tooltip: "No defects, normal sample"
      - name: "defective"
        tooltip: "Contains defects or anomalies"

  - annotation_type: "multiselect"
    name: "defect_types"
    description: "Select defect types (if defective)"
    labels:
      - name: "scratch"
        tooltip: "Surface scratches"
      - name: "dent"
        tooltip: "Dents or deformations"
      - name: "hole"
        tooltip: "Holes or punctures"
      - name: "crack"
        tooltip: "Cracks or fractures"
      - name: "contamination"
        tooltip: "Contamination or stains"
      - name: "color_defect"
        tooltip: "Color abnormalities"
      - name: "missing_part"
        tooltip: "Missing components"
      - name: "misalignment"
        tooltip: "Misaligned parts"

  - annotation_type: "text"
    name: "defect_location"
    description: "Describe the location of defects (e.g., 'top-left corner')"

interface_config:
  item_display_format: "<img src='{{text}}' style='max-width:100%; max-height:500px;'/><br/><small>Product type: {{context}}</small>"

output_annotation_format: "json"
output_annotation_dir: "annotations"

Sample Datasample-data.json

json
[
  {
    "id": "mvtec_001",
    "image_url": "https://upload.wikimedia.org/wikipedia/commons/thumb/8/8a/Banana-Single.jpg/1200px-Banana-Single.jpg",
    "product_type": "Example industrial product. Inspect for defects such as scratches, dents, holes, or contamination."
  },
  {
    "id": "mvtec_002",
    "image_url": "https://upload.wikimedia.org/wikipedia/commons/thumb/b/b1/VW_Lupo.jpg/1200px-VW_Lupo.jpg",
    "product_type": "Industrial component. Mark whether this is a normal (good) sample or contains defects."
  }
]

// ... and 1 more items

Get This Design

View on GitHub

Clone or download from the repository

Quick start:

git clone https://github.com/davidjurgens/potato-showcase.git
cd potato-showcase/image/specialized/mvtec-ad
potato start config.yaml

Dataset & paper

Bergmann, Fauser, Sattlegger & Steger, CVPR 2019

Official dataset ↗Read the paper ↗

Citation (BibTeX)

bibtex
@inproceedings{bergmann2019mvtec,
    title = {{MVTec AD} -- A Comprehensive Real-World Dataset for Unsupervised Anomaly Detection},
    author = {Bergmann, Paul and Fauser, Michael and Sattlegger, David and Steger, Carsten},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    pages = {9592--9600},
    year = {2019}
}

Details

Annotation Types

multiselectradiotext

Domain

Computer VisionIndustrial Inspection

Use Cases

Anomaly DetectionDefect DetectionQuality Control

Tags

mvtecanomaly-detectiondefectindustrialquality-controlcvpr2019

Found an issue or want to improve this design?

Open an Issue

Related Designs

CUB-200-2011: Fine-Grained Bird Species Dataset

CUB-200-2011 is a fine-grained bird image dataset of 11,788 photos across 200 species, each labeled with parts, attributes, and a bounding box (Wah et al., 2011). This Potato config reproduces the labeling task.

multiselectradio

FLAIR: French Land Cover from Aerospace Imagery

Land use and land cover classification from high-resolution aerial imagery. Annotators classify the primary land use category of aerial image patches and identify any secondary land uses present. Based on the FLAIR dataset from the French National Institute of Geographic and Forest Information (IGN).

multiselectradio

iWildCam Wildlife Detection & Classification

Camera trap image classification for wildlife monitoring (Beery et al., CVPR 2019). Classify wildlife species from camera trap images across diverse ecosystems worldwide.

multiselectradio