README.md

metadata

license: mit
task_categories:
  - image-classification
  - image-segmentation
tags:
  - anomaly-detection
  - cold-start

ArcAD Cold-Start Data Splits

Cold-start supervised data splits (JSON manifests) for MVTec-AD, VisA, Real-IAD, and MANTA, used by ArcAD: Anomaly-Rectified Calibration for Cold-Start Supervised Anomaly Detection (ECCV 2026).

• Repository: https://github.com/LGC-AD/ArcAD
• Paper: https://huggingface.co/papers/2607.02252

All paths use each dataset's original download structure — download the official datasets and the paths resolve directly. See the ArcAD repository for usage.

Split JSON format

Every <category>.json has the same schema:

{
  "meta":   { "dataset": "mvtec", "category": "bottle", "num_labeled": 69, "num_test": 223 },
  "labeled":[ { "image": "bottle/train/good/000.png",            "mask": "",                                              "label": 0, "anomaly_class": "good" },
              { "image": "bottle/test/broken_large/005.png",     "mask": "bottle/ground_truth/broken_large/005_mask.png", "label": 1, "anomaly_class": "broken_large" } ],
  "test":   [ ... ]
}

• All paths are relative to the dataset root (the --data_path argument) and use each dataset's original download layout.
• mask is "" for normal samples (no mask file).
• label: 0 = normal, 1 = anomaly.
• anomaly_class: "good" for normals; the defect sub-folder name (e.g. broken_large) for MVTec, "anomaly" for VisA / Real-IAD / MANTA.

The total number of labeled samples matches the cold-start protocol (e.g. MVTec-AD: 1089 normals + 121 anomalies; Real-IAD: 10940 normals + 1216 anomalies).

Expected on-disk layout

The JSON paths resolve against the official download structure of each dataset. Point --data_path at the root shown below:

MVTec-AD

It contains over 5000 high-resolution images divided into fifteen different object and texture categories.

<data_path>/bottle/
    train/good/*.png
    test/good/*.png
    test/<defect_type>/*.png            # e.g. broken_large, broken_small, contamination, ...
    ground_truth/<defect_type>/<name>_mask.png

VisA

It contains 12 subsets corresponding to 12 different objects. There are 10,821 images with 9,621 normal and 1,200 anomalous samples.

<data_path>/candle/
    Data/Images/Normal/*.JPG
    Data/Images/Anomaly/*.JPG
    Data/Masks/Anomaly/*.png

Real-IAD

A large-scale challenging industrial AD dataset, containing 30 classes with totally 151,050 images.

<data_path>/realiad_1024/<category>/<image>      # image_path from realiad_jsons/sup/<cat>.json
<data_path>/realiad_jsons/sup/<category>.json    # authoritative labeled/test split

MANTA

It contains 38 categories and over 130K object-level images.

<data_path>/MANTA_TINY_256_cropped/<category>/<image>
<data_path>/sup_cropped/<category>.json          # authoritative labeled/test split

Citation

If you find this work useful, please cite:

@article{han2026arcad,
  title   = {ArcAD: Anomaly-Rectified Calibration for Cold-Start Supervised Anomaly Detection},
  author  = {Han, Ningning and Fan, Lei and Guo, Jia and Cao, Yunkang and Su, Xiu and Cao, Feng and Di, Donglin and Su, Tonghua},
  journal = {arXiv preprint arXiv:2607.02252},
  year    = {2026}
}