MohammedEltoum/industrial-anomaly-playground

Industrial Anomaly Detection Playground 🏭

An interactive playground for industrial anomaly detection. Two methods run side by side on the same image so you can compare memorization-based and zero-shot detection:

Method	What it does	Needs
Memory bank (PatchCore-style)	DINOv2 patch features + nearest-neighbor distance against a bank of normal reference images; per-patch distances form the anomaly heatmap, and the image threshold is auto-calibrated on held-out normal references	~8 normal reference images of the object type
Zero-shot (WinCLIP-style)	CLIP with a compositional prompt ensemble ("flawless/damaged {object}") scored on the full image and a 3×3 grid of overlapping crops	nothing — no references at all

Features: 15 MVTec AD object/texture categories, clickable sample images (normal and defective), your own uploads, tunable thresholds, and heatmap overlays for both methods at once.

Run locally

pip install -r requirements.txt gradio
python app.py

The first launch downloads facebook/dinov2-base (350 MB) and openai/clip-vit-base-patch16 (600 MB) and takes 1–2 minutes. The first use of each category downloads 12 sample images (2 MB) and encodes the reference bank (~20 s on CPU), then caches it in memory.

Host it as a Hugging Face Space

The app is Space-ready. Create a Gradio Space and add these three files (app.py, requirements.txt, and this README): the Space needs only gradio from the SDK — the rest comes from requirements.txt. It runs on free CPU tier (models load in RAM, inference is a few seconds per image).

Data and attribution

• Sample images are MVTec AD test images (© MVTec Software GmbH, CC BY-NC-SA 4.0 — non-commercial use), fetched at runtime from the four-panel visualizations in abulhawa/industrial-component-anomaly-detection-models (top-left "Image" panel, 256×256). The category list and defect types match the MVTec AD dataset.
• Methods reimplemented after:
  • PatchCore — Roth et al., Towards Total Recall in Industrial Anomaly Detection, CVPR 2022, arXiv:2106.08265 (features swapped to DINOv2, as in AnomalyDINO, arXiv:2405.14529)
  • WinCLIP — Jeong et al., WinCLIP: Zero-/Few-Shot Anomaly Classification and Segmentation, CVPR 2023, arXiv:2303.14814
  • DINOv2 — Oquab et al., arXiv:2304.07193

Honest caveats: this is a demo, not a benchmark. Reference banks are built from ~8 normal images (not full training sets), thresholds are calibrated on 4 held-out normals, and the zero-shot map comes from 9 coarse crops rather than WinCLIP's full multi-scale sliding windows — expect somewhat lower accuracy than the papers' numbers, and treat both verdicts as exploratory.