Flownomaly recovers normal features for few-shot anomaly detection
Original title: Flownomaly: Single-Step Conditional-Mean Recovery for Few-Shot Anomaly Detection
In the 1-shot setting, image-level AUROC reached 97.8% on MVTec-AD and 95.3% on VisA.
Flownomaly treats few-shot industrial anomaly detection as recovering a normal counterpart in frozen DINOv2 feature space. A velocity field learned by conditional flow matching transports Gaussian noise toward normal patch tokens, and one recovery step yields the posterior mean without trajectory integration. A normal-only Jacobian criterion sets the operating time when no anomalous validation samples exist. On 1-, 2-, and 4-shot MVTec-AD and VisA, plus sample-shot Real-IAD, the paper reports state-of-the-art results against recent category-specific methods, including a 2.1-point image-level AUROC gain over the strongest VisA baseline.
- 1-shot pixel-level AUROC was 98.0% on MVTec-AD and 98.7% on VisA.
- Squared residuals support both localization and image-level detection.
- Operating time is chosen from normal supports alone.
- Code is released at the authors' GitHub repository.
- Products
- Flownomaly
- AI models
- DINOv2
From the original article
1 min read at arxiv.orgFew-shot industrial anomaly detection requires modeling category-specific normality from only a handful of defect-free images. Common approaches tie inference to stored exemplars, impose rigid linear structure, or risk reproducing anomalous content. Flownomaly instead casts the problem as normal-counterpart recovery in frozen DINOv2 feature space.
By Bernardo Perrone Ribeiro · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.