Research map / Label-efficient and robust learning / OOD and anomaly detection
MoECLIP: Patch-Specialized Experts for Zero-shot Anomaly Detection
Jun Yeong Park, JunYoung Seo, Minji Kang, Yu Rang Park
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 |
|---|---|---|---|
| add MoE patch-specialized LoRA experts with routing to frozen CLIP | CLIP · ZSAD avg image-level | AUROC: 85.5 -> 89.6 vs AA-CLIP | Improved |
| add MoE patch-specialized LoRA experts | CLIP · ZSAD avg pixel-level | AUROC: 93.2 -> 94.3 | Improved |
| add FOFS (frozen orthogonal feature separation) (ablation) | MoECLIP · 4-dataset average | image AUROC: 93.9 -> 95.3 | Improved |
| add ETF loss (ablation) | MoECLIP · 4-dataset average | image AUROC: 94.1 -> 95.3 | Improved |
| add both FOFS and ETF loss over naive MoE (ablation) | MoECLIP · 4-dataset average | pixel AUROC: 91.2 -> 92.2 | Improved |
| add depth-wise adapter (ablation) | MoECLIP · 4-dataset average | AUROC: 93.6 -> 95.3 image | Improved |
| add PAA (ablation) | MoECLIP · ColonDB pixel-level | AUROC: 81.9 -> 85.4 | Improved |
| add PAA (ablation) | MoECLIP · MVTec-AD pixel-level | AUROC: 92.1 -> 92.5 | Improved |
| add FOFS and ETF loss to reduce inter-expert similarity (ablation) | MoE experts · MVTec-AD expert cosine similarity | cosine similarity: 0.45 -> 0.02 (experts 1-2) | Improved |
| increase number of experts beyond 4 (ablation) | MoECLIP · MVTec-AD | AUROC | No clear effect |
| add CLIP adaptation (naive MoE LoRA) over vanilla CLIP (ablation) | CLIP · 4-dataset average | AUROC: 40.6 -> 91.2 pixel | Improved |
| use MoECLIP on Liver CT vs Bayes-PFL/AnomalyCLIP | MoECLIP · Liver CT image-level | AUROC: 68.2 -> 74.0 | Improved |
Closest papers by meaning
- AnomalyCLIP: Object-agnostic Prompt Learning for Zero-shot Anomaly Detection (ICLR 2024 · similarity 0.90)
- AdaptCLIP: Adapting CLIP for Universal Visual Anomaly Detection (AAAI 2026 · similarity 0.90)
- AA-CLIP: Enhancing Zero-Shot Anomaly Detection via Anomaly-Aware CLIP (CVPR 2025 · similarity 0.89)
- UniVAD: A Training-free Unified Model for Few-shot Visual Anomaly Detection (CVPR 2025 · similarity 0.87)
- Towards Zero-Shot Anomaly Detection and Reasoning with Multimodal Large Language Models (CVPR 2025 · similarity 0.87)
- MuSc: Zero-Shot Industrial Anomaly Classification and Segmentation with Mutual Scoring of the Unlabeled Images (ICLR 2024 · 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.
