Research map / Detection and segmentation / Open-set and few-shot detection
Triad: Empowering LMM-based Anomaly Detection with Expert-guided Region-of-Interest Tokenizer and Manufacturing Process
Yuanze Li, Shihao Yuan, Haolin Wang, Qizhang Li, Ming Liu, Chen Xu, Guangming Shi, Wangmeng Zuo
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 manufacturing process text to general-purpose LMMs | GPT-4o, Qwen2-VL, LLaVA-OneVision · MVTec-AD 0-shot | accuracy: GPT-4o 82.2% -> 67.9%; Qwen2-VL 7B 84.4% -> 61.1% | Hurt |
| add manufacturing process text to Triad | Triad-ov 7B · MVTec-AD 0-shot | accuracy: 91.2% -> 92.6% | Improved |
| train LLaVA-1.6-7B with Triad (InstructIAD, CoT-M, EG-RoI) | LLaVA-1.6-7B · MVTec-AD 0-shot | accuracy: 76.9% -> 85.0% | Improved |
| add InstructIAD fine-tuning (ablation) | LLaVA-1.6-mistral-7B · MVTec-AD 0-shot | accuracy: 76.9% -> 78.8% | Improved |
| add CoT-M data organization (ablation) | Triad with EG-RoI · MVTec-AD 0-shot + MFG process | accuracy: 83.9% -> 87.5% | Improved |
| add CoT-M data organization (ablation) | Triad with EG-RoI · MVTec-AD 0-shot (no MFG) | accuracy: 85.4% -> 85.0% | No clear effect |
| add EG-RoI tokenizer (ablation) | Triad (InstructIAD + CoT-M) · MVTec-AD 0-shot | accuracy: 79.8% -> 85.0% | Improved |
| replace AnyRes with EG-RoI (ablation) | LLaVA-1.6-mistral-7B fine-tuned · MVTec-AD small/medium/large defects | defect detection rate: small 90.9% -> 95.5%; large 65.7% -> 81.8%; normal 82.9% -> 72.8% | Mixed |
| use a higher-quality vision expert (MuSc vs April-GAN) (ablation) | EG-RoI · MVTec-AD 0-shot | accuracy: 83.0% -> 85.0% (base); P-AUROC 87.6% -> 97.3% | Improved |
| supply random boxes instead of expert regions (ablation) | EG-RoI · MVTec-AD 0-shot + MFG | accuracy: 87.5% -> 83.4% | Hurt |
| vary source of manufacturing process (Web, LLM, GPT) (ablation) | Triad-ov 7B · MVTec-AD 0-shot | accuracy: 91.2% -> 91.8% (Web), 92.6% (LLM), 92.2% (GPT) | Improved |
| provide one normal reference image (1-shot) | general-purpose LMMs · MVTec-AD | accuracy: GPT-4o 82.2% -> 77.6%; Qwen2-VL 72B 87.1% -> 92.0% | Mixed |
Closest papers by meaning
- Real-IAD D3: A Real-World 2D/Pseudo-3D/3D Dataset for Industrial Anomaly Detection (CVPR 2025 · similarity 0.89)
- AnomalyGPT: Detecting Industrial Anomalies Using Large Vision-Language Models (AAAI 2024 · similarity 0.89)
- MMAD: A Comprehensive Benchmark for Multimodal Large Language Models in Industrial Anomaly Detection (ICLR 2025 · similarity 0.88)
- 3CAD: A Large-Scale Real-World 3C Product Dataset for Unsupervised Anomaly Detection (AAAI 2025 · 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 0.88)
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.
