---
license: apache-2.0
---

# Visual Anomaly Detection under Complex View-Illumination Interplay: A Large-Scale Benchmark  
🌐 [Hugging Face Dataset](https://huggingface.co/datasets/ChengYuQi99/M2AD)  


> 📚 [**Paper**](https://arxiv.org/abs/2505.10996) • 🏠 [**Homepage**](https://hustcyq.github.io/M2AD/)  
> by [Yunkang Cao*](https://caoyunkang.github.io/), [Yuqi Cheng*](https://hustcyq.github.io/), [Xiaohao Xu](), [Yiheng Zhang](), [Yihan Sun](), [Yuxiang Tan](), [Yuxin Zhang](), [Weiming Shen](https://scholar.google.com/citations?user=FuSHsx4AAAAJ&hl=en),  


## 🚀 Updates  
We're committed to open science! Here's our progress:  
* **2025/05/19**: 📄 Paper released on [ArXiv](https://arxiv.org/abs/2505.10996).  
* **2025/05/16**: 🌐 Dataset homepage launched.  
* **2025/05/24**: 🧪 Code release for benchmark evaluation!  [code](https://github.com/hustCYQ/M2AD)


## 📊 Introduction  
Visual Anomaly Detection (VAD) systems often fail in the real world due to sensitivity to **viewpoint-illumination interplay**—complex interactions that distort defect visibility. Existing benchmarks overlook this challenge.  

Introducing **M2AD** (Multi-View Multi-Illumination Anomaly Detection), a large-scale benchmark designed to rigorously test VAD robustness under these conditions:  
- **119,880 high-resolution images** across 10 categories, 999 specimens, 12 views, and 10 illuminations (120 configurations).  
- **Two evaluation protocols**:  
  - 🔄 **M2AD-Synergy**: Tests multi-configuration information fusion.  
  - 🧪 **M2AD-Invariant**: Measures single-image robustness to view-illumination variations.  
- Key finding: SOTA VAD methods struggle significantly on M2AD, highlighting the critical need for robust solutions.