Paper presented at WACV 2025!
Our PhD Student Simon Damm presented the work “AnomalyDINO: Boosting Few-Shot Anomaly Detection with DINOv2” at the IEEE/CVF Winter Conference on Application in Computer Vision (WACV 2025) in Tucson, AZ, USA (Feb 28 – Mar 4).
The paper, a collboration between Simon Damm, Mike Laskiewicz, Asja Fischer and Johannes Lederer was accepted as an oral presentation at WACV 2025, and proposes a vision-only approach for few-shot anomaly detection.
The starting point from this research was the question “Do we need language for visual few-shot anomaly detection?” — motivated by previous SOTA techniques which all utilize language and vision, e.g., in a CLIP-style approach (common features space for visual and textual inputs) or by snythezising anomalies with Stable Diffusion (requiring text prompts describing normal and anomalous samples).
To answer this question, we revisit the established patch-level deep nearest neighbor paradigm, and propose AnomalyDINO — specifically designed for the few-shot regime utilizing the strong patch-level features from DINOv2. The proposed pipeline encompasses simple augmentations and robust zero-shot masking (discarding irrelevant background regions) to to ensure that diverse, but only relevant features are considered.
AnomalyDINO achieves new SOTA performance in the few-shot regime on MVTec-AD and VisA, two popular benchmark datasets for industrial anomaly detection, while singificantly reducing the average inference time compared to it’s closest competitors.
Check out the paper (CV open access, arXiv) and code.