Visual Anomaly Detection under Complex View-Illumination Interplay: A Large-Scale Benchmark

🌐 Hugging Face Dataset

📚 Paper • 🏠 Homepage
by Yunkang Cao*, Yuqi Cheng*, Xiaohao Xu, Yiheng Zhang, Yihan Sun, Yuxiang Tan, Yuxin Zhang, Weiming Shen,

🚀 Updates

We're committed to open science! Here's our progress:

• 2025/05/19: 📄 Paper released on ArXiv.
• 2025/05/16: 🌐 Dataset homepage launched.
• 2025/05/24: 🧪 Code release for benchmark evaluation!

📊 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.

🔍 Overview of M2AD

📷 Data Collection Process

10 Anomaly Categories

🛠️ Getting Started

📦 Environment Setup

Clone and install dependencies:

git clone https://github.com/hustCYQ/M2AD.git && cd M2AD  
conda create --name M2AD_env python=3.9 -y  
conda activate M2AD_env  

# Core dependencies  
pip install timm==0.8.15dev0 opencv-python==4.9.0.80 numpy==1.26  
pip install thop seaborn mmselfsup pandas transformers imgaug tensorboard  

# Optional extras  
pip install git+https://gitcode.com/gh_mirrors/cl/CLIP  
conda install -c conda-forge accimage  
pip install mmdet==2.25.3 adeval torch==2.1.2 --index-url https://download.pytorch.org/whl/cu118  

📂 Dataset Preparation

Our dataset is hosted on Hugging Face. You can also download via:

Dataset	Google Drive	Baidu Drive (🔑: soon!)	Resolution
M2AD-1024	Download	Download	1024x1024
M2AD-256	Download	Download	256x256

Data Structure:

M2AD_1024/  
├── Bird/  
│   ├── Good/       # Normal samples  
│   ├── GT/         # Ground-truth anomaly masks  
│   └── NG/         # Anomalous samples  
├── Car/  
... (10 categories total)  

Update DATA_ROOT and DATA_SUBDIR in /data/dataset_info to match your path.

🚀 Train & Evaluate

We provide ready-to-use scripts for 5 SOTA methods:

# Example: Train CDO on M2AD-Synergy  
CUDA_VISIBLE_DEVICES=0 python run_dataset.py --cfg_path configs/benchmark/cdo/cdo_100e.py -m train -s M2AD-Synergy  

# Evaluate on M2AD-Invariant  
CUDA_VISIBLE_DEVICES=0 python run_dataset.py --cfg_path configs/benchmark/msflow/msflow_100e.py -m test -s M2AD-Invariant  

Key Arguments:

• -s M2AD-Synergy: Evaluate multi-view/illumination fusion.
• -s M2AD-Invariant: Test single-image robustness.

Note: If needed, you can download dinov2_vitb14_reg4_pretrain.pth here.

🧩 Trainer Framework

Our modular trainer (extended from ADer) simplifies VAD development:

• Core components: forward(), compute_loss(), compute_anomaly_scores().
• Example implementation: trainer/cdo_trainer.py and model/cdo.py.
• Build new methods with minimal code changes!

📊 Main Results

1. Performance on M2AD (256x256)

2. Performance on M2AD (512x512)

🙏 Acknowledgements

Grateful to these projects for inspiration:

• 🌟 ADer – Anomaly detection framework.
• 🎨 EyeCandies – Project page.

📖 Citation

If M2AD aids your research, please cite:

@article{M2AD,  
  title={Visual Anomaly Detection under Complex View-Illumination Interplay: A Large-Scale Benchmark},  
  author={Cao, Yunkang and Cheng, Yuqi and Xu, Xiaohao and Zhang, Yiheng and Sun, Yihan and Tan, Yuxiang and Zhang, Yuxin and Huang, Xiaonan and Shen, Weiming},  
  journal={arXiv preprint arXiv:2505.10996},  
  year={2025}  
}  

Contact

If you have any questions about our work, please do not hesitate to contact [email protected].