README.md

---
license: cc-by-nc-4.0
task_categories:
- image-classification
- object-detection
- image-segmentation
language:
- en
tags:
- anomaly-detection
- industrial-inspection
- academic-only
- arxiv:2511.20169
pretty_name: ADNet
size_categories:
- 100K<n<1M
---

# ADNet

ADNet is a large-scale, multi-domain benchmark for visual anomaly detection and localization. It contains 196,294 RGB images from 380 real-world categories across five application domains: Electronics, Industry, Agrifood, Infrastructure, and Medical.

ADNet standardizes data collected from 49 publicly available anomaly-detection datasets into a unified MVTec-style structure. The benchmark supports normal-only training, image-level anomaly detection, pixel-level anomaly localization, and large-scale multi-class anomaly-detection research.

For further details, see the paper:

> **ADNet: A Large-Scale and Extensible Multi-Domain Benchmark for Anomaly Detection Across 380 Real-World Categories**  
> Hai Ling, Jia Guo, Zhulin Tao, Yunkang Cao, Donglin Di, Hongyan Xu, Xiu Su, Yang Song, and Lei Fan.  
> [arXiv:2511.20169](https://arxiv.org/abs/2511.20169)

## Access and Licensing

ADNet is released for **academic research purposes only**.

The dataset is distributed under the [CC BY-NC 4.0 License](https://creativecommons.org/licenses/by-nc/4.0/) (Attribution-NonCommercial 4.0 International). Commercial use is prohibited.

To request access, please apply using an institutional email address, such as an `.edu` or `.ac.cn` address.

- Requests from personal email addresses or commercial domains may be rejected.
- Please clearly state your institution and intended research use in the application form.
- Access requests are reviewed manually. Please allow 1–2 working days for approval.

## Dataset Overview

ADNet contains:

- **380 object categories**
- **5 application domains**
- **196,294 RGB images**
- **116,192 normal training images**
- **80,102 test images**
- **19,791 normal test images**
- **60,311 anomalous test images**
- **60,311 corresponding pixel-level anomaly masks**

The dataset is organized into five domain-level archives rather than separate archives for every category:

- `Agrifood.zip`
- `Electronics.zip`
- `Industry.zip`
- `Infrastructure.zip`
- `Medical.zip`

## Dataset Statistics

| Domain | Categories | Normal training images | Normal test images | Anomalous test images | Total images |
|:--|--:|--:|--:|--:|--:|
| Agrifood | 69 | 19,504 | 2,860 | 13,041 | 35,405 |
| Electronics | 77 | 22,886 | 3,989 | 15,105 | 41,980 |
| Industry | 154 | 44,278 | 7,680 | 20,983 | 72,941 |
| Infrastructure | 52 | 18,732 | 3,297 | 6,810 | 28,839 |
| Medical | 28 | 10,792 | 1,965 | 4,372 | 17,129 |
| **Total** | **380** | **116,192** | **19,791** | **60,311** | **196,294** |

Training sets contain only anomaly-free images. Each test set contains both normal and anomalous images, with pixel-level masks provided for anomalous samples.

## Dataset Structure

After extraction, ADNet follows a domain-category hierarchy with a unified MVTec-style organization:

```text
ADNet/
├── Agrifood/
│   └── <category>/
├── Electronics/
│   └── <category>/
├── Industry/
│   └── <category>/
├── Infrastructure/
│   └── <category>/
└── Medical/
    └── <category>/
```

Each category uses the following structure:

```text
<domain>/
└── <category>/
    ├── train/
    │   └── good/
    │       └── <normal training images>
    ├── test/
    │   ├── good/
    │   │   └── <normal test images>
    │   └── <defect_type>/
    │       └── <anomalous test images>
    └── ground_truth/
        └── <defect_type>/
            └── <pixel-level anomaly masks>
```

A concrete example is:

```text
ADNet/
└── Industry/
    └── screw_bag/
        ├── train/
        │   └── good/
        │       ├── 026.png
        │       ├── 077.png
        │       └── 078.png
        ├── test/
        │   ├── good/
        │   ├── logical_anomalies/
        │   │   └── 007.png
        │   └── structural_anomalies/
        │       └── 058.png
        └── ground_truth/
            ├── logical_anomalies/
            │   └── 007.png
            └── structural_anomalies/
                └── 058.png
```

For an anomalous test image, the corresponding mask has the same defect-type directory and filename:

```text
test/<defect_type>/<filename>
ground_truth/<defect_type>/<filename>
```

Normal training and test images do not have corresponding anomaly masks.

Image extensions may vary across the source datasets. Users should not assume that every image is a `.png` file or that mask filenames contain a `_mask` suffix.

## Supported Tasks

ADNet supports the following primary tasks:

- **Image-level anomaly detection:** determine whether a test image is normal or anomalous.
- **Pixel-level anomaly localization:** identify anomalous regions using the provided segmentation masks.
- **One-for-one anomaly detection:** train and evaluate one model for each category.
- **Multi-class anomaly detection:** jointly train a unified model using multiple categories or all 380 categories.
- **Cross-domain anomaly detection:** study transfer and generalization across the five application domains.