PatchCore Industrial Anomaly Detection State-of-the-art anomaly detection for industrial inspection achieving up to 99.6% AUROC on the MVTec AD benchmark using pretrained feature extractors and coreset subsampling
Quick start Get started with PatchCore in minutes
Install dependencies
Clone the repository and install required packages: git clone https://github.com/amazon-science/patchcore-inspection.git cd patchcore-inspection pip install -r requirements.txt
Download MVTec AD dataset
Download the MVTec AD benchmark dataset and organize it with the expected structure: # Download from https://www.mvtec.com/company/research/datasets/mvtec-ad # Organize as: mvtec/bottle/{train,test,ground_truth}
The dataset should contain 15 subdatasets: bottle, cable, capsule, carpet, grid, hazelnut, leather, metal_nut, pill, screw, tile, toothbrush, transistor, wood, and zipper.
Train your first model
Train PatchCore on a single category: env PYTHONPATH=src python bin/run_patchcore.py \ --gpu 0 --seed 0 --save_patchcore_model \ results patch_core -b wideresnet50 -le layer2 -le layer3 \ --faiss_on_gpu --pretrain_embed_dimension 1024 \ --target_embed_dimension 1024 --anomaly_scorer_num_nn 1 \ --patchsize 3 sampler -p 0.1 approx_greedy_coreset \ dataset --resize 256 --imagesize 224 -d bottle mvtec /path/to/mvtec
Evaluate the model
Load and evaluate your trained model: env PYTHONPATH=src python bin/load_and_evaluate_patchcore.py \ --gpu 0 --seed 0 results/evaluation \ patch_core_loader -p results/models/mvtec_bottle --faiss_on_gpu \ dataset --resize 256 --imagesize 224 -d bottle mvtec /path/to/mvtec
Image-level AUROC: 99.2% Pixel-level AUROC: 98.1% PRO Score: 94.4%
Key features Everything you need for industrial anomaly detection
State-of-the-art accuracy Achieve up to 99.6% image-level AUROC and 98.4% pixel-level AUROC on MVTec AD
Memory efficient Coreset subsampling reduces memory footprint while maintaining detection accuracy
50+ backbone networks Choose from ResNet, WideResNet, EfficientNet, Vision Transformers, and more
Ensemble support Combine multiple backbones and feature layers for even better performance
FAISS acceleration GPU-accelerated nearest neighbor search for fast inference
Pretrained models Ready-to-use models for all 15 MVTec AD categories
Explore by topic Deep dive into PatchCore capabilities
Ready to start detecting anomalies? Get up and running with PatchCore in minutes. Train your first model or explore our pretrained models for the MVTec AD benchmark.
