Research map / Label-efficient and robust learning / OOD and anomaly detection
SubspaceAD: Training-Free Few-Shot Anomaly Detection via Subspace Modeling
Camile Lendering, Erkut Akdag, Egor Bondarau
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| average DINOv2 layers 22-28 instead of the last layer only (ablation) | SubspaceAD · VisA (4-shot) | image AUROC: 89.1 -> 94.7 | Improved |
| average DINOv2 layers 22-28 instead of final 7 layers (ablation) | SubspaceAD · VisA (4-shot) | image AUROC: 91.9 -> 94.7 | Improved |
| average middle layers instead of concatenating them (ablation) | SubspaceAD · MVTec-AD (4-shot) | image AUROC: 98.6 -> 97.9 | Hurt |
| use middle-layer mean-pooling instead of final-7 mean-pooling (ablation) | SubspaceAD · MVTec-AD (4-shot) | image AUROC: 98.2 -> 97.9 | Hurt |
| keep tau = 0.99 instead of using the full feature space (tau = 1.00) (ablation) | SubspaceAD · MVTec-AD (1-shot) | image AUROC: 45.3 -> 97.1 | Improved |
| raise tau from 0.95 to 0.99 (ablation) | SubspaceAD · MVTec-AD (1-shot) | image AUROC: 98.2 -> 97.1 | Hurt |
| raise tau from 0.95 to 0.99 (ablation) | SubspaceAD · VisA (1-shot) | image AUROC: 92.2 -> 93.4 | Improved |
| use larger DINOv2 backbone (ablation) | SubspaceAD · MVTec-AD and VisA | anomaly detection metrics | Improved |
| increase input resolution beyond 448 px (ablation) | SubspaceAD · MVTec-AD and VisA | anomaly detection metrics | No clear effect |
| use DINOv3 backbones instead of DINOv2 (ablation) | SubspaceAD · MVTec-AD and VisA | anomaly detection metrics | Hurt |
| replace AnomalyDINO memory bank with PCA subspace | SubspaceAD · VisA (1-shot) | image AUROC: 87.4 -> 93.4 | Improved |
| replace AnomalyDINO memory bank with PCA subspace | SubspaceAD · VisA (4-shot) | PRO: 94.1 -> 93.8 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 2 on most shared results.
- PromptAD — PromptAD: Learning Prompts with only Normal Samples for Few-Shot Anomaly Detection (CVPR 2024) · better on 8 of 8 shared results
- MuSc — MuSc: Zero-Shot Industrial Anomaly Classification and Segmentation with Mutual Scoring of the Unlabeled Images (ICLR 2024) · better on 0 of 2 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- UniVAD: A Training-free Unified Model for Few-shot Visual Anomaly Detection (CVPR 2025 · similarity 0.89)
- Towards Training-free Anomaly Detection with Vision and Language Foundation Models (CVPR 2025 · similarity 0.87)
- Supervised Anomaly Detection for Complex Industrial Images (CVPR 2024 · similarity 0.87)
- MuSc: Zero-Shot Industrial Anomaly Classification and Segmentation with Mutual Scoring of the Unlabeled Images (ICLR 2024 · similarity 0.87)
- AnomalyGPT: Detecting Industrial Anomalies Using Large Vision-Language Models (AAAI 2024 · similarity 0.87)
- ResAD: A Simple Framework for Class Generalizable Anomaly Detection (NeurIPS 2024 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
