Research map / Detection and segmentation / Open-set and few-shot detection
Real-IAD: A Real-World Multi-View Dataset for Benchmarking Versatile Industrial 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.
Introduces Real-IAD, a 150K-image multi-view industrial anomaly detection dataset over 30 objects, and benchmarks popular IAD methods under standard unsupervised and a new fully unsupervised (noisy-training) setting.
Problem
Existing industrial anomaly detection datasets (MVTec AD, VisA) are saturated (>99% AUROC), small, single-view, and lack realistic settings with contaminated training data, limiting method comparison.
Approach
A large-scale real-world dataset with 30 object classes photographed from five viewpoints, pixel-level masks, sample-level metrics that fuse views, and a fully unsupervised IAD setting where training data contains 0-40% anomalous samples reflecting production yield rates above 60%.
Key results
- Recent IAD methods score about 97.9% mean I-AUROC on MVTec AD but about 85% (multi-view) to 87.6% (single-view) on Real-IAD, showing the dataset is harder.
- Real-IAD contains about 150K high-resolution images (2,000-5,000 px) across 30 object classes with five shooting angles each.
- In the fully unsupervised setting with noise ratios up to 0.4, most methods degrade severely, while PatchCore and SoftPatch remain the most robust.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Evaluate on Real-IAD instead of MVTec AD | IAD methods (reported mean) · I-AUROC | I-AUROC: 97.9 -> 87.6 | Hurt |
| Evaluate on Real-IAD (multi-view) instead of Real-IAD (single-view) | IAD methods (reported mean) · I-AUROC | I-AUROC: 87.6 ±2.9 -> 85.0 ±3.6 | Hurt |
| Add noisy anomalous training samples at α = 0.1 instead of α = 0.0 (ablation) | PatchCore · Real-IAD | S-AUROC: 94.4 -> 92.2 | Hurt |
| Add noisy anomalous training samples at α = 0.4 instead of α = 0.0 (ablation) | PatchCore · Real-IAD | S-AUROC: 94.4 -> 88.9 | Hurt |
| Add noisy anomalous training samples at α = 0.4 instead of α = 0.0 (ablation) | PatchCore · Real-IAD | I-AUROC: 91.3 -> 88.1 | Hurt |
| Add noisy anomalous training samples at α = 0.4 instead of α = 0.0 (ablation) | PatchCore · Real-IAD | P-PRO: 92.6 -> 92.4 | No clear effect |
| Add noisy anomalous training samples at α = 0.4 instead of α = 0.0 (ablation) | SoftPatch · Real-IAD | S-AUROC: 93.9 -> 90.6 | Hurt |
| Add noisy anomalous training samples at α = 0.4 instead of α = 0.0 (ablation) | SoftPatch · Real-IAD | I-AUROC: 91.4 -> 89.3 | Hurt |
| Add noisy anomalous training samples at α = 0.4 instead of α = 0.0 (ablation) | SoftPatch · Real-IAD | P-PRO: 92.1 -> 92.5 | Helped |
| Filter noisy features before constructing the memory bank | SoftPatch vs PatchCore · Real-IAD (FUIAD α = 0.4) | S-AUROC: 88.9 -> 90.6 | Helped |
Surprising findings
- Under the fully unsupervised setting, PaDiM, SimpleNet, DeSTSeg and RD degrade severely, while PatchCore stays relatively robust.
Limitations
- Only a few typical methods are benchmarked; zero-shot, few-shot and semi-supervised settings are not yet evaluated.
- Pixel-level localization on Real-IAD remains challenging, with lower P-AUPRO than MVTec AD.
- Multi-view setting abstracts away part flipping and bottom inspection used in real deployments.
Experimental setup
Datasets: Real-IAD MVTec AD VisA
Benchmarks: Real-IAD MVTec AD VisA
Metrics: I-AUROC P-AUPRO S-AUROC P-PRO
Baselines: PatchCore PaDiM CFlow SimpleNet DeSTSeg RD UniAD SoftPatch
Builds on: MVTec AD VisA SoftPatch
Closest papers by meaning
- Real-IAD D3: A Real-World 2D/Pseudo-3D/3D Dataset for Industrial Anomaly Detection (CVPR 2025 · similarity 0.92)
- Supervised Anomaly Detection for Complex Industrial Images (CVPR 2024 · similarity 0.90)
- AnomalyGPT: Detecting Industrial Anomalies Using Large Vision-Language Models (AAAI 2024 · similarity 0.89)
- 3CAD: A Large-Scale Real-World 3C Product Dataset for Unsupervised Anomaly Detection (AAAI 2025 · similarity 0.88)
- AnomalyNCD: Towards Novel Anomaly Class Discovery in Industrial Scenarios (CVPR 2025 · similarity 0.87)
- InCLAD: A Continual Learning Benchmark for Industrial Visual Anomaly Detection (NeurIPS 2026 · 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.