Research map / Label-efficient and robust learning / OOD and anomaly detection
Normal-Abnormal Guided Generalist Anomaly Detection
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes normal-abnormal guided generalist anomaly detection (NAGL), using one abnormal plus normal references via residual mining and anomaly feature learning attention modules on a frozen ViT-S encoder.
Problem
Generalist anomaly detection methods use only normal references, ignoring the small number of anomalous samples often available in practice; naive use of abnormal references causes false activations.
Approach
Model residuals between abnormal and normal references (Residual Mining) to learn transferable residual proxies, then map them onto query residuals (Anomaly Feature Learning) to produce anomaly proxies that are combined with a normal-reference nearest-neighbour score map.
Key results
- On MVTecAD with N1+A1, image AUROC 95.8 and pixel AUROC 96.6, beating the best N4 baseline WinCLIP (95.2/96.2).
- Ablation: naively merging normal and abnormal maps drops mean AUROC from 91.1 (normal only) to 88.4, while the full NAGL reaches 94.6.
- Efficiency: 24.4M parameters and 17.1 FPS, versus ResAD 59.2M/7.8 FPS and InCTRL 117.5M/1.2 FPS.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add one abnormal reference and process normal-abnormal residuals with NAGL rather than using normal-reference NN only (ablation) | Generalist anomaly detection · MVTecAD | Image AUROC: 93.2 -> 95.8; +2.6 | Helped |
| Add one abnormal reference and process normal-abnormal residuals with NAGL rather than using normal-reference NN only (ablation) | Generalist anomaly detection · VisA | Image AUROC: 81.5 -> 88.5; +7.0 | Helped |
| Naively add abnormal-reference NN scores to normal-reference NN scores without NAGL processing (ablation) | Generalist anomaly detection · MVTecAD and VisA | Image AUROC: 93.2 -> 90.7; 81.5 -> 77.2 | Hurt |
| Apply NAGL residual processing instead of naively merging normal- and abnormal-reference NN results (ablation) | Generalist anomaly detection · MVTecAD and VisA | Mean AUROC: 88.4% -> 94.6%; 6.2 | Helped |
| Increase abnormal references from one to two (ablation) | NAGL · MVTecAD and VisA | Mean AUROC: + 0.8; + 0.6; + 0.7 | Helped |
| Increase normal references from one to two to four while retaining one abnormal reference (ablation) | NAGL · VisA | Image F1-max: 85.4% to 86.8% and 87.8% | Helped |
| Increase normal references from one to two to four while retaining one abnormal reference (ablation) | NAGL · VisA | Pixel F1-max: 41.8% to 43.3% and reaches 44.0% | Helped |
| Use NAGL with one abnormal reference rather than the strongest reported baseline in the stated settings | Generalist anomaly detection · MVTecAD, N^1+A^1 vs N^1 baseline | Image AUROC: 95.8% vs 94.6% | Helped |
| Use NAGL with one abnormal reference rather than ResAD using four normal references | Generalist anomaly detection · VisA, N^2+A^1 vs N^4 | Image AUROC: 89.8% vs 89.3% | Helped |
| Use NAGL with two normal and one abnormal reference rather than InCTRL with two normal references | Generalist anomaly detection · BraTS | Image AUROC: 82.1% vs 74.6% | Helped |
Surprising findings
- Naively adding abnormal reference scores degrades performance below normal-only detection, due to false activations.
- Abnormal-reference-only KNN gives much weaker results (mean AUROC 74.2) than normal-only.
Limitations
- Evaluated only on image data; video and time series not explored.
- Only one abnormal reference used in main experiments; more abnormal references only tested in appendix.
Experimental setup
Datasets: MVTecAD VisA
Benchmarks: MVTecAD VisA BraTS MVTec3D MVTecLOCO BTAD MPDD
Metrics: AUROC AP F1-max PRO
Baselines: SPADE PaDiM PatchCore WinCLIP PromptAD InCTRL ResAD RegAD MVFA
Builds on: PatchCore InCTRL ResAD ViT
Closest papers by meaning
- Toward Generalist Anomaly Detection via In-context Residual Learning with Few-shot Sample Prompts (CVPR 2024 · similarity 0.91)
- ResAD: A Simple Framework for Class Generalizable Anomaly Detection (NeurIPS 2024 · similarity 0.89)
- Beyond Normal References: Discriminative Few-Shot Anomaly Detection (NeurIPS 2026 · similarity 0.89)
- Open-Set Graph Anomaly Detection via Normal Structure Regularisation (ICLR 2025 · similarity 0.88)
- AnomalyNCD: Towards Novel Anomaly Class Discovery in Industrial Scenarios (CVPR 2025 · similarity 0.88)
- One-to-Normal: Anomaly Personalization for Few-shot 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.