SubspaceAD: Training-Free Few-Shot Anomaly Detection via Subspace Modeling
Camile Lendering, Erkut Akdag, Egor Bondarev · 2026
Read SubspaceAD: Training-Free Few-Shot Anomaly Detection via Subspace Modeling by Camile Lendering, Erkut Akdag, Egor Bondarev (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2602-23013.txt.
Description
Camile Lendering and two co-authors ask whether few-shot anomaly detection needs its usual complexity. Industrial inspection often has only a few normal images per category, and recent methods lean on memory banks, auxiliary datasets, or multi-modal tuning of vision-language models. SubspaceAD instead works training-free in two simple stages.
Patch-level features come from a small set of normal images through a frozen DINOv2 backbone, then a principal component analysis model fits the low-dimensional subspace of normal variation. At inference, anomalies score by reconstruction residual against that subspace, giving interpretable, statistically grounded scores. One-shot results reach about ninety-seven percent image- and pixel-level AUROC on MVTec-AD. On VisA they reach about ninety-three and ninety-eight percent, surpassing prior state-of-the-art results.
Catalog record
| CREATOR | Camile Lendering, Erkut Akdag, Egor Bondarev |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 47 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2602-23013