mvtec-anomaly

Industrial Anomaly Detection: PatchCore vs EfficientAD on MVTec AD

Compares two SOTA unsupervised anomaly detection methods on the MVTec AD benchmark — a memory-based approach (PatchCore) and a knowledge-distillation approach (EfficientAD) — with a Gradio demo for interactive defect visualisation.

Structure

mvtec-anomaly/
├── app.py                    # Gradio demo
├── data/
│   ├── mvtec_dataset.py      # Dataset loader 
│   └── download_mvtec.sh     # Download instructions
├── models/
│   ├── patchcore.py          # PatchCore (CVPR 2022)
│   └── efficientad.py        # EfficientAD (WACV 2024)
├── scripts/
│   ├── train.py              # Train + evaluate, auto-generates plots
│   └── plot_results.py       # Heatmap grids + comparison charts
├── utils/
│   ├── paths.py              # Central path config via .env
│   └── metrics.py            # AUROC, AP, PRO score
└── results/                  # Generated plots (gitignored)

Setup

python -m venv venv && source venv/bin/activate
pip install -r requirements.txt

Download MVTec AD (free, requires registration):

bash data/download_mvtec.sh

Train & Evaluate

# Single category, both models
python scripts/train.py --category bottle

# All 15 categories
python scripts/train.py --category all

# Single model
python scripts/train.py --category bottle --model patchcore

Gradio Demo

python app.py          # local
python app.py --share  # public link

Methods

Model	Paper	Key Idea	Inference Speed
PatchCore	CVPR 2022	Memory bank of normal patch features	Medium
EfficientAD	WACV 2024	Teacher-Student knowledge distillation	Fast

Metrics

Metric	Description
Image AUROC	Discriminates normal vs anomalous images
Pixel AUROC	Localises defects at pixel level
PRO Score	Per-Region Overlap — primary MVTec metric