Research map / Vision-language and multimodal / CLIP and zero-shot
DeMuS: Learning Decoupled Matching and Scoring for Batch Zero-Shot Industrial Anomaly Detection
Zhengyang Zhao, Hailong Sun, Binhang Qi, Hongrui Yu, Zhongchi Wang, Hang Xu
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use DINOv3 backbone instead of CLIP (training-free) (ablation) | mutual scoring · MVTec AD | pixel AUROC / AUPRO: +2.88% pixel AUROC, +3.18% AUPRO | Improved |
| train only the matching adapter (no scoring adapter) (ablation) | DeMuS · MVTec AD and VisA average | image AUROC: 85.6 -> 82.6 | Hurt |
| add matching adapter to scoring-only variant (decoupling) (ablation) | DeMuS · VisA | image AP: -2.49% drop in AP on VisA without it | Improved |
| use multi-scale convolutional adapter instead of MLP bottleneck (ablation) | DeMuS · MVTec AD | image AUROC: -4.91% drop on MVTec AD with MLP | Improved |
| add MuSc-style multi-scale aggregation to MLP adapter (ablation) | MLP adapter · MVTec AD and VisA average | image AUROC: 89.2 -> 89.0 | No clear effect |
| use alignment-based correspondence oracle instead of average-pooled NN matching (ablation) | DeMuS · MVTec AD and VisA average | image AUROC: 89.5 -> 92.3 | Improved |
| add pose augmentation in training (ablation) | DeMuS · MVTec AD and VisA average | image AUROC: 89.9 -> 92.3 | Improved |
| add pose augmentation in training (ablation) | DeMuS · MVTec AD (pose variation) | image AUROC retention: -4.3% vs -3.0% | Improved |
| add pose augmentation in training (ablation) | DeMuS · VisA (pose variation) | image AUROC retention: -14.7% vs -6.8% | Improved |
| increase batch size from 2 to 4 (ablation) | DeMuS · MVTec AD, VisA, Real-IAD average | image AUROC: 92.28 at B=4; 24.27 ms per image | Improved |
| increase batch size beyond 4 (ablation) | DeMuS · average | image AUROC / latency: gains saturate while latency rises | No clear effect |
| increase top-k nearest neighbors from 1 to 5 (ablation) | DeMuS · average | image-level vs pixel-level: image 92.90 -> 93.30, pixel 95.11 -> 94.62 | Mixed |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- AdaptCLIP — AdaptCLIP: Adapting CLIP for Universal Visual Anomaly Detection (AAAI 2026) · better on 4 of 4 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- MuSc-V2: Zero-Shot Multimodal Industrial Anomaly Classification and Segmentation With Mutual Scoring of Unlabeled Samples (TPAMI 2026 · similarity 0.88)
- MuSc: Zero-Shot Industrial Anomaly Classification and Segmentation with Mutual Scoring of the Unlabeled Images (ICLR 2024 · similarity 0.88)
- InCLAD: A Continual Learning Benchmark for Industrial Visual Anomaly Detection (NeurIPS 2026 · similarity 0.87)
- AnomalyGPT: Detecting Industrial Anomalies Using Large Vision-Language Models (AAAI 2024 · similarity 0.87)
- Exploring Intrinsic Normal Prototypes within a Single Image for Universal Anomaly Detection (CVPR 2025 · similarity 0.87)
- Supervised Anomaly Detection for Complex Industrial Images (CVPR 2024 · 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.
