---
language: en
license: cc-by-nc-sa-4.0
tags:
- anomaly-detection
- computer-vision
- industrial-inspection
- patchcore
- mvtec-ad
datasets:
- foersben/mvtec-ad
metrics:
- roc_auc
- f1
library_name: pytorch
pipeline_tag: image-feature-extraction
---

# Industrial Anomaly Detection: PatchCore (MVTec AD Bottle)

## Model Overview
- **Architecture:** PatchCore (Roth et al., CVPR 2022)
- **Backbone:** Frozen `Wide-ResNet-50-2` (Layers 2 & 3 concatenation, D = 1536 channels)
- **Memory Bank:** 10% Greedy k-Center minimax coreset subsampling (16,386 prototypes)
- **Evaluation Split:** Official MVTec AD Bottle hold-out test set (83 images: 20 good, 63 defect with ground-truth masks)

## Empirical Benchmark Results (Tesla T4 GPU)
| Metric | Target | Result |
| :--- | :---: | :---: |
| **Image-level AUROC** | >= 0.980 | **1.0000 (100% Perfect Separation)** |
| **Pixel-level AUROC** | >= 0.960 | **0.9848** |
| **AUPRO (FPR <= 0.30)** | >= 0.920 | **0.9376** |
| **Mean Latency** | < 100 ms | **26.31 ms** |
| **Median Latency (p50)** | — | **26.25 ms** |
| **95th Percentile Latency (p95)** | < 80 ms | **27.51 ms** |

## Threshold Methodology Disclosure
- **Oracle In-Sample Threshold:** Threshold tau = 17.6255 yields F1 = 1.0000 on the 83 test images. Disclosed as an empirical upper bound.
- **Operational Calibration (Recommended):** In production lines without defect samples, establish threshold on a held-out nominal validation split (`train/good`) at the 99th percentile (target FAR <= 1.0%).

## Hardware & Limitations
- **Hardware:** Evaluated on NVIDIA Tesla T4 GPU (15.0 GB VRAM, Turing architecture).
- **Domain Scope:** Specialized for clear rigid glassware bottles. Requires a separate nominal prototype memory bank for opaque steel surfaces or textiles.