Skip to main content

​
About MVTec AD

The MVTec Anomaly Detection (MVTec AD) dataset is the standard benchmark for industrial anomaly detection. It contains high-resolution images of 15 different object and texture categories with various types of defects.
Dataset Statistics
  • 15 product categories
  • 5,354 images total (3,629 training + 1,725 testing)
  • Training images: only defect-free (“good”) samples
  • Test images: defect-free and various anomaly types
  • Ground truth masks for pixel-level evaluation

​
Download the Dataset

1

Visit MVTec Website

Go to the official MVTec AD dataset page:

MVTec AD Downloadhttps://www.mvtec.com/company/research/datasets/mvtec-ad
Fill out the download form to access the dataset.
2

Download and Extract

Download the complete dataset (approximately 4.9 GB compressed).
Extract Dataset
# Extract the downloaded archive
tar -xvf mvtec_anomaly_detection.tar.xz
# or
unzip mvtec_anomaly_detection.zip
3

Verify Directory Structure

Ensure the extracted folder follows the correct structure (see below).

​
Required Directory Structure

The dataset must be organized with the following structure for PatchCore to load it correctly:
Directory Tree
mvtec/
├── bottle/
│   ├── ground_truth/
│   │   ├── broken_large/
│   │   ├── broken_small/
│   │   ├── contamination/
│   │   └── ...
│   ├── test/
│   │   ├── good/
│   │   ├── broken_large/
│   │   ├── broken_small/
│   │   ├── contamination/
│   │   └── ...
│   └── train/
│       └── good/
├── cable/
│   ├── ground_truth/
│   ├── test/
│   └── train/
├── capsule/
├── carpet/
├── grid/
├── hazelnut/
├── leather/
├── metal_nut/
├── pill/
├── screw/
├── tile/
├── toothbrush/
├── transistor/
├── wood/
└── zipper/
The directory structure must match this layout exactly. PatchCore expects:
  • train/good/ for training images
  • test/ with subdirectories for each defect type
  • ground_truth/ with pixel-level masks for defects

​
Dataset Categories

MVTec AD includes 15 categories across objects and textures:
Objects (rigid items with consistent appearance):
  • bottle - Glass bottles
  • cable - Power cables
  • capsule - Pharmaceutical capsules
  • hazelnut - Hazelnuts
  • metal_nut - Metal nuts
  • pill - Pharmaceutical pills
  • screw - Screws
  • toothbrush - Toothbrushes
  • transistor - Electronic transistors
  • zipper - Zippers

​
Verify Installation

Check that your dataset is properly structured: 
# Navigate to dataset directory
cd /path/to/mvtec

# Count categories (should be 15)
ls -1 | wc -l

# Check a specific category structure
ls -R bottle/

​
Configure Dataset Path

Once the dataset is ready, you’ll reference it in training commands: 
# Set environment variable
export MVTEC_DATA=/path/to/mvtec

# Use in training commands
python bin/run_patchcore.py [...] dataset [...] mvtec $MVTEC_DATA

​
Dataset Loading Process

When you run training, PatchCore loads the dataset as follows:
1

Locate Category

For each specified category (e.g., bottle), PatchCore looks for:
{data_path}/bottle/train/good/*.png
2

Apply Transformations

Each image is:
  1. Resized to --resize (default: 256)
  2. Center-cropped to --imagesize (default: 224)
  3. Converted to tensor
  4. Normalized with ImageNet mean/std
Normalization
IMAGENET_MEAN = [0.485, 0.456, 0.406]
IMAGENET_STD = [0.229, 0.224, 0.225]
3

Create DataLoader

Images are batched with:
  • batch_size: Default 2 (configurable)
  • num_workers: Default 8 (parallel loading)
  • shuffle: False (deterministic order)

​
Training on Subset of Categories

You don’t have to train on all 15 categories. Specify only the ones you need:
Train Single Category
# Train only on bottle category
python bin/run_patchcore.py --gpu 0 --seed 0 results \
  patch_core -b wideresnet50 -le layer2 -le layer3 \
  --pretrain_embed_dimension 1024 --target_embed_dimension 1024 \
  --anomaly_scorer_num_nn 1 --patchsize 3 --faiss_on_gpu \
  sampler -p 0.1 approx_greedy_coreset \
  dataset --resize 256 --imagesize 224 -d bottle mvtec /path/to/mvtec
Train Multiple Categories
# Train on specific categories
python bin/run_patchcore.py --gpu 0 --seed 0 results \
  patch_core [...] sampler [...] \
  dataset --resize 256 --imagesize 224 \
    -d bottle -d cable -d capsule \
    mvtec /path/to/mvtec

​
Common Issues

Problem: Dataset path is incorrect or structure doesn’t match expected format.Solution:
  1. Verify the path exists: ls /path/to/mvtec
  2. Check category exists: ls /path/to/mvtec/bottle
  3. Verify train folder: ls /path/to/mvtec/bottle/train/good
Problem: Training images not in correct location.Solution:
  • All training images must be in {category}/train/good/
  • Check for .png or .jpg files in that directory
  • Verify file permissions (readable)
Problem: Mask files missing or in wrong location.Solution:
  • Masks should be in {category}/ground_truth/{defect_type}/
  • Mask filenames must match test image filenames
  • Only defect images have masks (good samples don’t need masks)

​
Next Steps

Train Single Model

Start training with WideResNet50 backbone

Configuration Guide

Learn about all training parameters

Build docs developers (and LLMs) love

Get started for freeTalk to us