Research map / Detection and segmentation / Open-set and few-shot detection
3CAD: A Large-Scale Real-World 3C Product Dataset for Unsupervised 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 3CAD, a large real-world pixel-annotated anomaly detection dataset from 3C production lines, and a coarse-to-fine method CFRG combining heterogeneous distillation, recovery guidance, and segmentation.
Problem
Existing industrial anomaly detection datasets (MVTec-AD, VisA, Real-IAD) have few defect samples, mostly artificial defects, and are saturated; no large real-world dataset exists for 3C product defects, and small subtle defects are hard to localize.
Approach
A coarse-to-fine paradigm where a heterogeneous WideResNet50-teacher/EfficientNet-b0-student distillation model gives coarse localization, a recovery network's features weight the anomaly maps, and a segmentation head refines the final mask.
Key results
- CFRG achieves 93.4% P-AUROC, 86.5% I-AUROC, 82.0% P-PRO and 17.6% AP on 3CAD, versus 92.3% P-AUROC for RD (mean across classes in the paper's tables).
- On MVTec-AD, CFRG reports 98.4% P-AUROC, 98.4% I-AUROC, 95.6% P-PRO and 73.4% AP, averaged across categories.
- Most popular methods reach around 99% on MVTec-AD but drop by over 10 points on 3CAD, with AP often below 10%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use CFRG instead of CRAD | 3CAD · 3CAD | I-AUROC: 79.3 -> 86.5 | Helped |
| Use CFRG instead of RD++ | 3CAD · 3CAD | P-PRO: 77.2 -> 82.0 | Helped |
| Add the recovery branch (ablation) | CFRG · 3CAD | P-AUROC: +6.8% | Helped |
| Add the recovery branch (ablation) | CFRG · 3CAD | AP | Hurt |
| Add the segmentation branch (ablation) | CFRG · 3CAD | AP: +9.1% | Helped |
| Keep the abnormal-feature distance push in distillation (ablation) | CFRG · 3CAD | I-AUROC: 0.6% | Helped |
| Add the recovery-branch auxiliary weight to the segmentation input (ablation) | CFRG · 3CAD | AP: 1.5% | Helped |
| Use a heterogeneous EfficientNet-b0 student instead of a WideResNet50 student (ablation) | CFRG · 3CAD | P-PRO: 1.1% | Helped |
| Set loss weights to 1, 1, 1 instead of 1, 1, 1/10 (ablation) | CFRG · 3CAD | AP: 14.2 -> 17.6 | Helped |
| Set loss weights to 1, 1, 1 instead of 1, 1/10, 1 (ablation) | CFRG · 3CAD | P-AUROC: 92.7 -> 93.4 | Helped |
Surprising findings
- Removing the recovery branch drops P-PRO sharply (59.8 vs 82.0), while AP is slightly higher without it (19.7 vs 17.6).
- Using a heterogeneous student (EfficientNet-b0) instead of a WideResNet50 student improves results only modestly (e.g., +0.5 P-AUROC).
Limitations
- Absolute AP remains low (17.6% at best), indicating difficulty in fine pixel-level localization.
- Dataset source is a single company's production lines, which may limit generality.
- Some reported numbers are inconsistent across tables (e.g., mean values and metric labels), so comparisons need care.
Experimental setup
Datasets: 3CAD MVTec-AD DTD
Benchmarks: 3CAD MVTec-AD
Metrics: I-AUROC P-AUROC P-PRO AP
Baselines: PaDiM FastFlow RD RD++ SimpleNet DREAM DeSTSeg RealNet UniAD CRAD
Builds on: Reverse Distillation (RD) DRAEM RD++ FastFlow
Closest papers by meaning
- Real-IAD D3: A Real-World 2D/Pseudo-3D/3D Dataset for Industrial Anomaly Detection (CVPR 2025 · similarity 0.89)
- Supervised Anomaly Detection for Complex Industrial Images (CVPR 2024 · similarity 0.88)
- Real-IAD: A Real-World Multi-View Dataset for Benchmarking Versatile Industrial Anomaly Detection (CVPR 2024 · similarity 0.88)
- Multimodal Industrial Anomaly Detection by Crossmodal Feature Mapping (CVPR 2024 · similarity 0.87)
- Towards Generalizable 3D Anomaly Detection via Relational Inconsistency Modeling (NeurIPS 2026 · similarity 0.87)
- AnomalyNCD: Towards Novel Anomaly Class Discovery in Industrial Scenarios (CVPR 2025 · 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.