CVPR 2026
AIMS Group, Department of Electrical Engineering, Eindhoven University of Technology
SubspaceAD is a training-free method for few-shot industrial anomaly detection. It extracts patch-level features from a small set of anomaly-free images using a frozen DINOv2 backbone and fits a PCA model to estimate a low-dimensional subspace that captures normal appearance variation. At inference, anomalies are detected and localized via reconstruction residuals with respect to this subspace, yielding interpretable and statistically grounded anomaly scores. SubspaceAD reports state-of-the-art performance across one-shot and few-shot settings without memory banks, auxiliary datasets, prompt tuning, or any form of training.
Overview of SubspaceAD. (Fitting) Patch features from normal images are extracted by a frozen DINOv2 model and a PCA subspace is fitted. (Inference) Test features are projected onto the normal subspace and reconstruction errors yield an anomaly map.
Qualitative anomaly localization examples.
@inproceedings{lendering2026subspacead,
title={SubspaceAD: Training-Free Few-Shot Anomaly Detection via Subspace Modeling},
author={Lendering, Camile and Akdag, Erkut and Bondarev, Egor},
booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
pages={28557--28566},
year={2026}
}