Research map / Label-efficient and robust learning / OOD and anomaly detection

Beyond Normal References: Discriminative Few-Shot Anomaly Detection

NeurIPS 2026 · Label-efficient and robust learning › OOD and anomaly detection · approach: Unified multi-class anomaly detection · 0 citations (OpenAlex)

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.

Introduces discriminative few-shot anomaly detection with both normal and anomalous references and proposes IDEAL, which learns intrinsic deviation vectors from denoised residual deviations to score anomalies.

Problem

Generalist few-shot anomaly detection uses only normal references, ignoring discriminative information in a few anomalous references; naively matching to anomalous references overfits to seen anomaly types.

Approach

Normal Variation Eraser removes PCA-estimated local normal variation subspace from residual deviations between anomalous and normal patch features; an Intrinsic Deviation Encoder uses cross-attention with learnable queries to learn M mutually orthogonal deviation vectors; anomaly score is based on query-to-normal deviation preserved after projection onto these vectors.

Key results

• With N1A1 setting, IDEAL reaches 96.3 image AUROC / 96.8 pixel AUROC on MVTecAD, above the best normal-only and NAGL baselines (NAGL 95.1 / 96.1).
• Under the hard setting on MPDD with N1A1, IDEAL attains 74.5 image AUROC versus NAGL's 63.7, showing better generalization to unseen anomalies.
• IDEAL has 23.83M parameters and runs at 21.9 FPS, about 14x faster than InCTRL and 1.3x faster than NAGL.

Surprising findings

• Directly matching queries to anomalous references (anomaly-only or combined) yields noisy, spurious activations, so anomalous references cannot be used naively as matching templates.

Limitations

• Performance on hard medical datasets such as BraTS and AITEX stays relatively low (e.g., AITEX image AUROC 77.6 at N2A1)
• Relies on a fixed single anomaly-type reference set per dataset, so coverage of anomaly types is limited

Experimental setup

Datasets: MVTecAD VisA AITEX MPDD BTAD BraTS Liver RESC

Benchmarks: MVTecAD VisA AITEX MPDD BTAD BraTS

Metrics: AUROC F1-Max PRO AP

Baselines: RegAD PromptAD WinCLIP InCTRL ResAD FoundAD DRA AHL NAGL

Builds on: NAGL InCTRL ResAD

Closest papers by meaning

• One-to-Normal: Anomaly Personalization for Few-shot Anomaly Detection (NeurIPS 2024 · similarity 0.89)
• Normal-Abnormal Guided Generalist Anomaly Detection (NeurIPS 2025 · similarity 0.89)
• ResAD: A Simple Framework for Class Generalizable Anomaly Detection (NeurIPS 2024 · similarity 0.87)
• Toward Generalist Anomaly Detection via In-context Residual Learning with Few-shot Sample Prompts (CVPR 2024 · similarity 0.87)
• UniVAD: A Training-free Unified Model for Few-shot Visual Anomaly Detection (CVPR 2025 · similarity 0.87)
• Dinomaly: The Less Is More Philosophy in Multi-Class Unsupervised Anomaly Detection (CVPR 2025 · similarity 0.87)