title	MVTec Anomaly Benchmark
emoji	🔍
colorFrom	blue
colorTo	purple
sdk	gradio
sdk_version	6.3.0
app_file	app.py
pinned	false
license	mit

MVTec Anomaly Benchmark

A comprehensive benchmark for anomaly detection models on the MVTec AD dataset using Anomalib.

🚀 Try the Live Demo on Hugging Face Spaces

🎯 Features

• Multiple Models: PatchCore, EfficientAD, FastFlow, STFPM, PaDiM
• Full Benchmark: Train and evaluate on all 15 MVTec categories
• Interactive Demo: Gradio UI for real-time anomaly detection
• Sample Image Gallery: Browse and select sample images from MVTec dataset with automatic category detection
• Draw Defects: Draw artificial defects on images and see how models detect them
• Model Comparison: Compare multiple models side-by-side on the same image
• Easy Configuration: YAML-based model configs

📦 Installation

# Clone the repository
git clone https://github.com/YOUR_USERNAME/mvtec-anomaly-benchmark.git
cd mvtec-anomaly-benchmark

# Create virtual environment (optional but recommended)
python -m venv venv
source venv/bin/activate  # Linux/Mac
# or: venv\Scripts\activate  # Windows

# Install dependencies
pip install -r requirements.txt

📥 Download Dataset

Interactive Downloader:

python scripts/download_mvtec.py

The script features an interactive menu where you can choose between:

1. Hugging Face (Recommended - Fast): Downloads from micguida1/mvtech_anomaly_detection. No login required.
2. HTTP Mirror (Fallback): Downloads from the original public mirror (~5GB, slower).

The dataset will be automatically extracted to data/MVTecAD/.

⬇️ Download Pre-trained Checkpoints

Checkpoints are hosted on HuggingFace Hub to keep this repository lightweight.

# Download all checkpoints
python scripts/download_checkpoints.py

# Download specific model/category
python scripts/download_checkpoints.py --model patchcore --category bottle

Note: Update HF_REPO_ID in scripts/download_checkpoints.py with your HuggingFace repository.

🚀 Usage

Training

# Train PatchCore on bottle category (default)
python train.py

# Train specific model on specific category
python train.py --model patchcore --category bottle

# Train all models on all categories
python train.py --model all --category all

# Train EfficientAD on hazelnut
python train.py --model efficientad --category hazelnut

Inference

# Run inference on a single image
python inference.py --image_path path/to/image.png --model patchcore --category bottle

Gradio Demo

python app.py

The demo will be available at http://localhost:7860.

Demo Features

• 📤 Upload Image: Upload any image and analyze it for anomalies
• ✏️ Draw Defects: Load a sample image and draw artificial defects to test detection
• 🔄 Compare Models: Compare multiple models side-by-side on the same image
• 📚 Learn: Educational content about each anomaly detection model
• 📊 Metrics: View detailed performance metrics for each model

Sample Image Gallery: Each tab includes a gallery of sample images from the MVTec dataset. Click on any image to load it and the category will be automatically selected.

📁 Project Structure

mvtec-anomaly-benchmark/
├── app.py                 # Gradio demo entry point
├── train.py               # Training script
├── inference.py           # Inference script
├── requirements.txt       # Python dependencies
├── configs/               # Model configurations (YAML)
│   ├── patchcore.yaml
│   ├── efficientad.yaml
│   └── ...
├── core/                  # Core module (config, models, utils)
├── gradio_ui/             # Gradio UI components
├── scripts/               # Utility scripts
│   ├── download_mvtec.py        # Download dataset
│   └── download_checkpoints.py  # Download from HF Hub
├── data/                  # Dataset directory
│   └── MVTecAD/
├── results/               # Training results & checkpoints (on HF Hub)
└── output/                # Inference outputs

☁️ Upload Checkpoints to HuggingFace Hub

After training, upload your checkpoints to HuggingFace Hub:

# Login to HuggingFace
huggingface-cli login

# Create a new model repository
huggingface-cli repo create mvtec-anomaly-checkpoints --type model

# Upload the results folder
huggingface-cli upload YOUR_USERNAME/mvtec-anomaly-checkpoints results/ .

Then update HF_REPO_ID in scripts/download_checkpoints.py.

🤖 Available Models

Model	Paper	Description
PatchCore	CVPR 2022	Memory bank with coreset subsampling
EfficientAD	WACV 2024	Lightweight student-teacher
FastFlow	arXiv 2021	Normalizing flows
STFPM	arXiv 2021	Student-Teacher Feature Pyramid
PaDiM	ICPR 2021	Patch Distribution Modeling

📊 MVTec AD Categories

The benchmark covers all 15 categories:

Textures	Objects
Carpet, Grid, Leather, Tile, Wood	Bottle, Cable, Capsule, Hazelnut, Metal Nut, Pill, Screw, Toothbrush, Transistor, Zipper

📄 License

This project is licensed under the MIT License - see the LICENSE file for details.

🙏 Acknowledgments

• Anomalib - Anomaly detection library
• MVTec AD - Dataset

💻 Computational Environment

All experiments were conducted on a cloud machine rented via Lightning.ai with the following specifications:

Component	Specification
CPU	Intel® Xeon® Platinum 8468 (16 vCPUs, 8 physical cores @ 2.1 GHz)
RAM	196 GB
GPU	NVIDIA H200 (141 GB HBM3 VRAM)

This high-performance setup enabled fast training and evaluation of all models across the entire MVTec AD dataset.