Research map / Label-efficient and robust learning / Zero-/few-shot and meta-learning
MuSc: Zero-Shot Industrial Anomaly Classification and Segmentation with Mutual Scoring of the Unlabeled Images
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 MuSc, a training-free zero-shot industrial anomaly classification and segmentation method that scores each test image's patches by mutual nearest-neighbor distances to other unlabeled test images, with multi-degree neighborhood aggregation and graph-based image-level re-scoring.
Problem
Zero-shot industrial anomaly classification (AC) and segmentation (AS) without normal training images or text prompts, exploiting the unlabeled test images themselves.
Approach
Normal patches find many similar patches in other unlabeled test images while anomalous patches find few; the method uses this via mutual scoring (min-distance to other images, interval averaging of lowest 30%), multi-degree local aggregation of ViT patch tokens, and a constrained kNN graph re-scoring (RsCIN) on class tokens.
Key results
- MuSc reaches 97.8% image AUROC and 97.3% pixel AUROC on MVTec AD in the zero-shot setting, with PRO rising from 72.7% to 93.8% over the prior best zero-shot method.
- On VisA, MuSc gets 92.8% classification AUROC (+14.7 over APRIL-GAN) and 98.8% segmentation AUROC.
- On the BTAD dataset, MuSc reaches 94.8% image AUROC in zero-shot, versus 72.3% for APRIL-GAN zero-shot.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use MuSc instead of the strongest reported zero-shot baseline | zero-shot industrial anomaly detection · MVTec AD | PRO-segm: 72.7% to 93.8% (+21.1%) | Helped |
| Use MuSc instead of the strongest reported zero-shot baseline | zero-shot industrial anomaly detection · VisA | AUROC-cls: +14.7% | Helped |
| Use aggregation degrees {1, 3, 5} rather than a single aggregation degree (ablation) | MuSc · MVTec AD | pixel-AUROC: 94.6 -> 97.3 | Helped |
| Use aggregation degrees {1, 3} instead of {1, 3, 5} (ablation) | MuSc · VisA | image-AUROC: 92.8 -> 93.7 | Helped |
| Average scores from all images instead of selecting the minimum 30% interval (ablation) | MSM · MVTec AD | image-AUROC: 97.8 -> 95.0 | Hurt |
| Replace 30% interval averaging with maximum scoring (ablation) | MSM · MVTec AD | image-AUROC: 97.8 -> 96.0 | Hurt |
| Add RsCIN (ablation) | MuSc · VisA | image-AUROC: 90 -> 92.8 | Helped |
| Add RsCIN (ablation) | MuSc · MVTec AD | F1-max: 96.6 -> 97.5 | Helped |
| Add aggregation degrees {7, 9, 13} to {1, 3, 5} (ablation) | MuSc LNAMD · MVTec AD | image-AUROC: 97.8 -> 95.7 | Hurt |
| Exclude the minimum 10% of scores from the selected 30% interval (ablation) | MSM · MVTec AD | image-AUROC: 97.8 -> 97.1 | Hurt |
Surprising findings
- Using the unlabeled test set alone outperforms several 4-shot and some full-shot one-class methods.
- Few-shot reference images give little improvement because the unlabeled test images already carry much normal information.
Limitations
- Inference time and memory grow with the number of test images, requiring subset splitting for large test sets.
- Sensitive to inconsistent orientations and scales because the frozen ViT backbone has only minor augmentation-based invariance.
- RsCIN requires a separate image-level feature backbone for one-class methods without class tokens.
Experimental setup
Datasets: MVTec AD VisA BTAD
Benchmarks: MVTec AD VisA BTAD
Metrics: AUROC AP F1-max PRO pixel AUROC pixel AP
Baselines: WinCLIP APRIL-GAN ACR RegAD PatchCore GraphCore
Builds on: CLIP PatchCore SSO
Closest papers by meaning
- AnomalyNCD: Towards Novel Anomaly Class Discovery in Industrial Scenarios (CVPR 2025 · similarity 0.87)
- Few Shot Part Segmentation Reveals Compositional Logic for Industrial Anomaly Detection (AAAI 2024 · similarity 0.87)
- UniVAD: A Training-free Unified Model for Few-shot Visual Anomaly Detection (CVPR 2025 · similarity 0.87)
- Supervised Anomaly Detection for Complex Industrial Images (CVPR 2024 · similarity 0.86)
- DZAD: Diffusion-based Zero-shot Anomaly Detection (AAAI 2025 · similarity 0.86)
- 3CAD: A Large-Scale Real-World 3C Product Dataset for Unsupervised Anomaly Detection (AAAI 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.