Research map / Label-efficient and robust learning / OOD and anomaly detection
Unveiling Multi-View Anomaly Detection: Intra-view Decoupling and Inter-view Fusion
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 IDIF, a single multi-view anomaly detection model that decouples view-common and view-specific features via a Consistency Bottleneck, fuses views with a learnable implicit 3D voxel, and uses View-wise Dropout to handle missing views.
Problem
Single-view anomaly detection misses anomalies in blind spots, and ensembles of single-view models ignore correlations between views; multi-view detection also breaks when some views are missing at test time.
Approach
Decouple each sample's multi-view features into view-common (via a variational bottleneck maximizing mutual information) and view-specific parts, fuse view-specific features through a learnable sample-independent 3D voxel prototype with cross-attention, and train with view-wise dropout plus masked-view cross attention so the model tolerates missing views.
Key results
- Achieves 97.0% S-AUROC and 98.9% P-AUROC on Real-IAD, outperforming the second-best method by 2.1% S-AUROC.
- Reaches 95.6% S-AUROC on MvTec 3D-AD and 94.2% on Eyecandies.
- Ablation: ensemble of DeSTSeg 94.0% -> single-model IVC 96.1% -> adding Consistency Bottleneck 97.0% S-AUROC; View-wise Dropout keeps 95.9% with one missing view and 94.0% with four.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use IDIF instead of SimpleNet | Real-IAD anomaly detection · Real-IAD (Average All) | S-AUROC (%): 94.9 -> 97.0 | Helped |
| use IDIF instead of PatchCore | Real-IAD anomaly localization · Real-IAD (Average All) | P-AUROC (%): 98.5 -> 98.9 | Helped |
| use IDIF instead of MCFM | MVTec 3D-AD anomaly detection · MVTec 3D-AD | S-AUROC (%): 95.4 -> 95.6 | Helped |
| use IDIF instead of MMRD | Eyecandies anomaly detection · Eyecandies | S-AUROC (%): 94.0 -> 94.2 | Helped |
| replace a DKD ensemble of single-view models with one multi-view model using 3D-convolution fusion (ablation) | Real-IAD multi-view anomaly detection · Real-IAD | Sample AUROC: 94.0 -> 95.8 | Helped |
| replace 3D-convolution fusion with Implicit Voxel Construction (IVC) (ablation) | Real-IAD multi-view anomaly detection · Real-IAD | Sample AUROC: 95.8 -> 96.1 | Helped |
| add Consistency Bottleneck (CB) before IVC fusion (ablation) | Real-IAD multi-view anomaly detection · Real-IAD | Sample AUROC: 96.1 -> 97.0 | Helped |
| add View-wise Dropout training | IDIF on full-view Real-IAD test data · Real-IAD | S-AUROC: 97.0% to 96.4% | Hurt |
| train with View-wise Dropout and test with one view missing | IDIF · Real-IAD, one randomly missing view | S-AUROC: 95.9% | Hurt |
| train with View-wise Dropout and test with two views missing | IDIF · Real-IAD, two randomly missing views | S-AUROC: 95.7% | Hurt |
| train with View-wise Dropout and test with three views missing | IDIF · Real-IAD, three randomly missing views | S-AUROC: 95.1% | Hurt |
| train with View-wise Dropout and test with four views missing | IDIF · Real-IAD, four randomly missing views | S-AUROC: 94.0% | Hurt |
Surprising findings
- View-wise Dropout slightly lowers full-view performance (97.0% to 96.4%) as a tradeoff for robustness to missing views.
Limitations
- Missing-view robustness costs about 0.6 points of full-view S-AUROC.
- MvTec 3D-AD and Eyecandies are converted into synthetic multi-view 2D data rather than native multi-view captures.
- Compute and hardware details are not stated.
Experimental setup
Datasets: Real-IAD MvTec 3D-AD Eyecandies DTD
Benchmarks: Real-IAD MvTec 3D-AD Eyecandies
Metrics: S-AUROC P-AUROC
Baselines: PaDim PatchCore RD UniAD DeSTSeg SimpleNet AST M3DM MMRD MCFM
Builds on: DeSTSeg Variational information bottleneck Spatial transformer networks Real-IAD
Closest papers by meaning
- AnomalyGPT: Detecting Industrial Anomalies Using Large Vision-Language Models (AAAI 2024 · similarity 0.85)
- Real-IAD: A Real-World Multi-View Dataset for Benchmarking Versatile Industrial Anomaly Detection (CVPR 2024 · similarity 0.85)
- Real-IAD D3: A Real-World 2D/Pseudo-3D/3D Dataset for Industrial Anomaly Detection (CVPR 2025 · similarity 0.84)
- Beyond Single-Modal Boundary: Cross-Modal Anomaly Detection through Visual Prototype and Harmonization (CVPR 2025 · similarity 0.84)
- Towards Training-free Anomaly Detection with Vision and Language Foundation Models (CVPR 2025 · similarity 0.84)
- Exploring Intrinsic Normal Prototypes within a Single Image for Universal Anomaly Detection (CVPR 2025 · similarity 0.84)
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.