Industrial anomaly-detection model that needs only good parts: memory bank of nominal patch features + nearest-neighbor matching reaches 99.1% image AUROC on MVTec AD — the reference method for cold-start visual defect detection.
Input
image
Output
Not specified
Industrial anomaly-detection model that needs only good parts: memory bank of nominal patch features + nearest-neighbor matching reaches 99.1% image AUROC on MVTec AD — the reference method for cold-start visual defect detection.
Save it, compare it or record that you have used it — with an account.
PatchCore (Roth et al., CVPR 2022, Amazon Science with University of Tübingen) detects and localizes visual defects using only defect-free training images. It stores locally-aware patch features from a pretrained backbone in a coreset-subsampled memory bank; at inference, patches far from their nearest nominal neighbor are flagged anomalous, yielding both an image-level score and a pixel-level anomaly map. It achieves up to 99.1% image-level AUROC on MVTec AD and stays strong in low-shot regimes (training from a handful of good images), while remaining CPU-deployable — properties that made it the default baseline for industrial QC. Reference code is Apache-2.0; a maintained implementation ships in Anomalib.
This product is part of the Birdwave Atlas community.
Contribute evaluations, share integrations and help improve compatibility data.
Discussions, evaluations, integrations and updates
Discussions and reviews are for members
Sign in to read what members say about PatchCore and to add your own evaluation, integration notes or spec fixes.
Sign in to join the discussion