README.md

Datasets Descriptions for Anomaly Detection

---

Both Semantic and Sensory AD

COCO-AD

• Download and extract COCO 2017 into data/coco.
• runpython data/gen_benchmark/coco.py to obtain data/coco/meta_20_x.json and val2017_mask_ad_20_x, where x=0,1,2,3 represent splits.
• Use DefaultAD in data/ad_dataset.py as the dataloader.

data
├── coco
    ├── annotations
        ├── instances_train2017.json
        ├── instances_val2017.json
    ├── train2017
    ├── val2017
    ├── meta_20_0.json
    ├── val2017_mask_ad_20_0
        ├── 000000000139.png
        ├── 000000000724.png

Sensory AD

MVTec AD

• Download and extract MVTec AD into data/mvtec.
• runpython data/gen_benchmark/mvtec.py to obtain data/mvtec/meta.json that matches standard DefaultAD in data/ad_dataset.py.

data
├── mvtec
    ├── meta.json
    ├── bottle
        ├── train
            └── good
                ├── 000.png
        ├── test
            ├── good
                ├── 000.png
            ├── anomaly1
                ├── 000.png
        └── ground_truth
            ├── anomaly1
                ├── 000.png

MVTec 3D-AD

• Download and extract MVTec 3D-AD into data/mvtec3d.
• runpython data/gen_benchmark/mvtec.py to obtain data/mvtec3d/meta.json that matches standard DefaultAD in data/ad_dataset.py.

VisA

• Download and extract VisA into data/visa.
• Refer to project page for data preparation, and run.
• runpython data/gen_benchmark/visa.py to obtain data/visa/meta.json that matches standard DefaultAD in data/ad_dataset.py.

Real-IAD

• Download and extract Real-IAD into data/realiad.

Uni-Medical

• Download and extract Uni-Medical into data/medical.
• Refer to ViTAD and BMAD for data declaration.

Semantic AD

Cifar

• Download and extract Cifar10 / Cifar100 into data/cifar
• Use CifarAD in data/ad_dataset.py as the dataloader, which covers three general settings for Cifar10 and one unified setting for Cifar100.

Dataset	Setting	Description
Cifar10	Unified	5 normals & 5 abnormals 5x5,000 for train & 5x1,000+5x1,000 for test
Cifar10	One-Class-Train	1 normal & 9 abnormals 1x5,000 for train & 1x1,000+1,000 for test
Cifar10	One-Class-Test	9 normals & 1 abnormal 9x5,000 for train & 9x1,000+6,000 for test
Cifar100	Unified	50 normals & 50 abnormals 50x500 for train & 50x100+50x100 for test

data
├── cifar
    ├── cifar-100-python
    ├── cifar-10-batches-py

Tiny-ImageNet-200

• Download and extract Tiny ImageNet into data/tiny-imagenet-200.
• runpython data/gen_benchmark/coco.py to obtain data/coco/meta_20_x.json and val2017_mask_ad_20_x, where x=0,1,2,3 represent splits.
• Use TinyINAD in data/ad_dataset.py as the dataloader.

data
├── cifar
    ├── cifar-100-python
    ├── cifar-10-batches-py

Dataset	Setting	Description
Cifar100	Unified	100 normals & 100 abnormals 100x500 for train & 100x50+100x50 for test