ReinAD: Towards Real-world Industrial Anomaly Detection with a Comprehensive Contrastive Dataset

Our dataset consists of a training set and a test set. All normal and anomaly images are in hdf5 format. In the mask annotations, pixels with a value of 0 represent normal regions, and pixels with a value of 1 represent anomaly regions.

The file structure of the training set and the test set are consistent, as follows:

dataset/
├── train/
│   ├── category1.h5
│   ├── category2.h5
│   └── ...
│
└── test/
    ├── category1.h5
    ├── category2.h5
    └── ...

The structure of the hdf5 file is as follows, where chunk_size = 100:

/ (root)
├── attrs
│   ├── split: "train"/"test"
│   └── category: category_name
│
├── Images
│   ├── Anomaly_0: [chunk_size, H, W, C]       # Anomaly images
│   ├── Anomaly_1: [chunk_size, H, W, C]
│   ├── ...
│   ├── Normal_0: [chunk_size, H, W, C]        # Normal images
│   ├── Normal_1: [chunk_size, H, W, C]
│   └── ...
│
└── Masks
    ├── Anomaly_0: [chunk_size, H, W]          # Pixel-level annotations for anomaly images
    ├── Anomaly_1: [chunk_size, H, W]
    └── ...