Industrial Anomaly Detection with DINOv3

Overview

A computer-vision project exploring how far you can push zero- and few-shot defect detection on industrial images using a frozen self-supervised backbone (DINOv3), with no fine-tuning on the target classes.

The benchmark is MVTec AD, the standard for industrial anomaly detection.

Approach

1. Embedding similarity at the image level. Compute DINOv3 features for “good” reference images, score test images by distance to that reference distribution.
2. Patch-level localization. Extract dense patch embeddings and compare each patch against its nearest neighbours in the reference set to produce a heatmap of where the anomaly likely is.
3. Few-shot extension. Use a handful of labeled defect examples to anchor the similarity scoring instead of relying purely on the “good” distribution.

Takeaway

Strong frozen features carry you surprisingly far on industrial data, without any class-specific training. But choosing the right similarity metric and the right pooling level matters as much as the backbone itself.