AnomalyNCD: Towards Novel Anomaly Class Discovery in Industrial Scenarios

Get Anomaly Maps: [Back to Catalogue]

Our AnomalyNCD could receive anomaly maps output from any anomaly detection (AD) method as input. To accurately implement the results in our paper, we provide download links to anomaly maps output by some typical anomaly detection methods below.

In order to be compatible with more custom anomaly detection methods, we provide the following code to save anomaly maps. Note that all datasets need to be pre-processed to the structure of the MVTec AD dataset.

def generate_anomaly_maps(image_paths, anomaly_maps, output_dir):
    """
    image_paths (list): List of image paths
    anomaly_maps (torch.Tensor): Anomaly maps from AD method
    output_dir (str): The directory that saves the anomaly maps
    """
    def normalization01(imgs):
        return (imgs - imgs.min()) / (imgs.max() - imgs.min())

    anomaly_maps_norm = normalization01(anomaly_maps)

    for i, path in enumerate(image_paths):
        anomaly_type = path.split('/')[-2]
        img_name = path.split('/')[-1]
        category = path.split('/')[-4]

        save_path = os.path.join(output_dir, category, anomaly_type)
        os.makedirs(save_path, exist_ok=True)
        save_path = os.path.join(save_path, img_name)
        anomaly_map = anomaly_maps_norm[i]*255
        cv2.imwrite(save_path, anomaly_map.astype(np.uint8))

We use MuSc as an example to describe how the above code can be used to save anomaly maps.

# Code from [MuSc](https://github.com/xrli-U/MuSc/blob/main/models/musc.py)

print('computing metrics...')
pr_sp = np.array(scores_cls)
gt_sp = np.array(gt_list)
gt_px = torch.cat(img_masks, dim=0).numpy().astype(np.int32)
pr_px = np.array(anomaly_maps)
image_metric, pixel_metric = compute_metrics(gt_sp, pr_sp, gt_px, pr_px)
auroc_sp, f1_sp, ap_sp = image_metric
auroc_px, f1_px, ap_px, aupro = pixel_metric

"""
Basically the code will provide the three parameters we need, 
you need to find image_paths and category in code.
"""

# our code to generate anomaly maps
save_dir = 'mvtec_ad_anomaly_map'
generate_anomaly_maps(image_path_list, pr_px, save_dir)

The anomaly maps on the MVTec AD dataset are stored according to the following structure.

data
|-- mvtec_ad_anomaly_map
|-----|-- bottle
|-----|-----|-- broken_large
|-----|-----|-----|-- 000.png
|-----|-----|-----|-- 001.png
|-----|-----|-----|-- ...
|-----|-----|-- broken_small
|-----|-----|-- ...
|-----|-- cable
|-----|-- ...

bash scripts/anomalyncd.sh

@inproceedings{huang2025anomalyncd,
  title={AnomalyNCD: Towards Novel Anomaly Class Discovery in Industrial Scenarios},
  author={Huang, Ziming and Li, Xurui and Liu, Haotian and Xue, Feng and Wang, Yuzhe and Zhou, Yu},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  year={2025}
}