🔔 News

• 2026-07: Code and data-split JSONs for MVTec-AD / VisA / Real-IAD / MANTA are released.
• ArcAD is accepted to ECCV 2026.

📖 Introduction

Deploying Industrial Anomaly Detection (IAD) in real manufacturing frequently hits a cold-start bottleneck: very few normal samples are available to represent the full normal distribution, and only a handful of anomalies are at hand. Under this regime, existing methods struggle to form a compact normal boundary and fail to exploit the rare supervised defect signal.

ArcAD (Anomaly-Rectified Cold-start AD) is a plug-and-play calibration framework built on top of reconstruction-based IAD baselines. Under data scarcity, it constructs a compact and discriminative normal boundary by combining hypersphere-based prototype modeling (SPM) with defect-guided contrastive calibration (DGC).

Extensive experiments on MVTec-AD, VisA, Real-IAD, and MANTA show that ArcAD clearly outperforms state-of-the-art supervised and unsupervised methods in both single-class and multi-class settings under cold-start conditions.

Figure 1: Overall framework of ArcAD.

📊 Results

Multi-class unified cold-start setting (all categories trained in a single model). Each cell reports I-AUROC / P-AUROC / P-F1-max (%) (Table 1 of the paper).

ArcAD is built upon Dinomaly as the reconstruction baseline. Per-category results are written to saved_results/<save_name>/results.csv after evaluation.

🛠️ Environment

conda create -n arcad python=3.8
conda activate arcad
pip install -r requirements.txt

Tested on a single NVIDIA RTX 3090 (24GB) with PyTorch 1.12 + CUDA 11.3. Key dependencies: torch, torchvision, timm, scikit-learn, opencv-python-headless, tabulate.

The frozen DINOv2-reg ViT-B/14 backbone weights should be placed at:

backbones/weights/dinov2_vitb14_reg4_pretrain.pth

(Download from the official DINOv2 release; the registers variant.)

📁 Data Preparation

ArcAD adopts a cold-start supervised split: a small labeled set (few normals + few anomalies, with masks) for training, and a test set for evaluation. The exact splits we used are released as open JSON files on 🤗 Hugging Face:

Dataset splits: https://huggingface.co/datasets/nnh1012/ArcAD_Cold-start_Data_Splits

Each split JSON is a manifest — a list of which images (and their masks) belong to the labeled / test sets. All paths are written against the original download structure of each dataset — just download the official datasets, set --data_path to the root, and the relative paths resolve directly, with no reorganization needed. The manifests do not contain the images themselves.

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:

<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

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

<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

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

🚀 Usage

Training and evaluation run in the same script (the model is evaluated on the test set every eval_freq iterations and at the end). Each dataset has its own entry-point script.

Step 1 — Generate prototypes (once per dataset)

Prototypes are K-means centers of the normal-sample bottleneck features, used to initialize SPM. A single generator handles all datasets via --dataset:

python gen_protos.py --dataset mvtec
python gen_protos.py --dataset visa
python gen_protos.py --dataset manta
python gen_protos.py --dataset realiad

This writes prototypes_init.pth into the dataset's init directory. The number of prototypes K defaults to 500 (override with --num_prototypes N). Set the GPU with GEN_DEV=cuda:N if needed.

Single-class setting: pass --separate_classes to emit one prototype file per category (prototypes_init_<category>.pth) instead of a single global file.

Step 2 — Train + evaluate

# MVTec-AD
python arcad_mvtec_uni.py \
    --data_path /path/to/mvtec_CD \
    --save_name arcad_mvtec

# VisA
python arcad_visa_uni.py \
    --data_path /path/to/VisA_CD \
    --save_name arcad_visa

# Real-IAD
python arcad_realiad_uni.py \
    --data_path /path/to/Real-IAD_CD \
    --save_name arcad_realiad

# MANTA
python arcad_manta_uni.py \
    --data_path /path/to/MANTA_CD \
    --save_name arcad_manta

Each script points to the prototype file generated in Step 1 by default; override the path with --proto_path if you placed it elsewhere.

Results, checkpoints, and logs are saved under saved_results/<save_name>/:

• results.csv — per-category + Mean metrics
• model.pth — final weights
• log.txt — training log

📂 Repository Structure

ArcAD/
├── arcad_{mvtec,visa,manta,realiad}_uni.py   # per-dataset train+eval entry points
├── gen_protos.py                              # unified prototype generator (--dataset)
├── dataset.py                                 # dataset classes (cold-start splits)
├── models/                                    # ViTill reconstruction backbone, SPM, DGC
├── backbones/                                 # frozen DINOv2 encoder loader
├── optimizers/                                # StableAdamW, cosine scheduler
├── utils.py                                   # metrics, logging
└── requirements.txt

🙏 Acknowledgements

This implementation is built upon Dinomaly. We also thank the authors of RD4AD and ReContrast for their excellent work, as well as the maintainers of MVTec-AD, VisA, Real-IAD, and MANTA.

📜 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}
}

Paper: https://arxiv.org/pdf/2607.02252