Research map / Label-efficient and robust learning / OOD and anomaly detection
Anomaly Heterogeneity Learning for Open-set Supervised 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 Anomaly Heterogeneity Learning (AHL), a generic OSAD framework that simulates heterogeneous anomaly distributions and collaboratively learns a unified abnormality model with a self-supervised importance estimator, plugged into DRA and DevNet.
Problem
Open-set supervised anomaly detection (OSAD) methods treat the few seen anomaly examples as a homogeneous distribution, which limits generalization to unseen anomaly classes that can come from very different distributions.
Approach
Generate many heterogeneous anomaly distributions by pairing normal clusters with randomly sampled seen anomalies and pseudo anomalies, train base OSAD models on support sets, and optimize a unified model on held-out query sets (surrogate open-set environment), weighting base models by a self-supervised sequential importance estimator.
Key results
- Under the general setting with ten anomaly examples, AHL (DRA) reaches 0.970 mean AUC on MVTec AD versus 0.966 for DRA, and improves DRA on Optical from 0.966 to 0.976.
- With one anomaly example, AHL (DRA) improves DRA on Optical from 0.880 to 0.888 and on Mastcam from 0.709 to 0.743 (AUC).
- AHL (DRA) is the best performer on AITEX and SDD among open-set detectors compared with the fully supervised FS-DRA on Table 4 and outperforms unsupervised KDAD and PatchCore on most datasets.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply AHL to DRA | DRA · AITEX, 10 anomaly examples (random) | AUC: 0.892±0.007 -> 0.925±0.013 | Helped |
| Apply AHL to DRA | DRA · SDD, 10 anomaly examples (random) | AUC: 0.990±0.000 -> 0.991±0.000 | Helped |
| Apply AHL to DRA | DRA · ELPV, 10 anomaly examples (random) | AUC: 0.843±0.002 -> 0.850±0.004 | Helped |
| Apply AHL to DRA | DRA · Optical, 10 anomaly examples (random) | AUC: 0.966±0.002 -> 0.976±0.004 | Helped |
| Apply AHL to DRA | DRA · Mastcam, 10 anomaly examples (random) | AUC: 0.849±0.003 -> 0.855±0.005 | Helped |
| Apply AHL to DRA | DRA · BrainMRI, 10 anomaly examples (random) | AUC: 0.971±0.001 -> 0.977±0.001 | Helped |
| Apply AHL to DevNet | DevNet · HeadCT, 10 anomaly examples (random) | AUC: 0.997±0.002 -> 0.999±0.003 | Helped |
| Apply AHL to DRA | DRA · Hyper-Kvasir, 10 anomaly examples (random) | AUC: 0.844±0.001 -> 0.880±0.003 | Helped |
| Apply AHL to DRA | DRA · MVTec AD (mean), 10 anomaly examples (random) | AUC: 0.966±0.002 -> 0.970±0.002 | Helped |
| Add HADG-generated heterogeneous distributions and CDL (ablation) | DRA · AITEX, general setting, 10 examples | AUC: 0.892±0.007 -> 0.916±0.004 | Helped |
| Replace full-data DRA with an ensemble trained using different random seeds (ablation) | DRA · AITEX, general setting, 10 examples | AUC: 0.892±0.007 -> 0.905±0.011 | Helped |
| Use random data subsets instead of HADG-generated subsets with CDL (ablation) | DRA · Mastcam, general setting, 10 examples | AUC: 0.852±0.001 -> 0.815±0.003 | Hurt |
Surprising findings
- Ensembling DRA with different random seeds gives only marginal or negative gains on several datasets, while the learned unified model improves consistently.
- AHL (DevNet) underperforms DevNet on some one-shot settings such as Optical (0.547 vs 0.513 baseline-level results are mixed), and cross-domain transfer drops on some target domains.
Limitations
- Cross-domain performance drops on some target domains, possibly due to bias from source-domain anomalies.
- Gains are smaller for weaker base models and for some datasets (e.g., AHL (DevNet) below DevNet on Optical in one-shot).
Experimental setup
Datasets: MVTec AD AITEX SDD ELPV Optical Mastcam HeadCT BrainMRI Hyper-Kvasir
Benchmarks: MVTec AD AITEX SDD ELPV Optical Mastcam HeadCT BrainMRI Hyper-Kvasir
Metrics: AUC
Baselines: SAOE MLEP FLOS DevNet DRA KDAD PatchCore FS-DRA
Builds on: DRA DevNet CutPaste DRAEM CutMix
Closest papers by meaning
- ResAD: A Simple Framework for Class Generalizable Anomaly Detection (NeurIPS 2024 · similarity 0.90)
- SetAD: Semi-Supervised Anomaly Learning in Contextual Sets (NeurIPS 2026 · similarity 0.89)
- Long-Tailed Anomaly Detection with Learnable Class Names (CVPR 2024 · similarity 0.88)
- Distribution Prototype Diffusion Learning for Open-set Supervised Anomaly Detection (CVPR 2025 · similarity 0.88)
- Normal-Abnormal Guided Generalist Anomaly Detection (NeurIPS 2025 · similarity 0.87)
- Towards Training-free Anomaly Detection with Vision and Language Foundation Models (CVPR 2025 · similarity 0.86)
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.