The authors of the MVTec AD: the MVTec Anomaly Detection dataset addressed the critical task of detecting anomalous structures within natural image data, a crucial aspect of computer vision applications. To facilitate the development of methods for unsupervised anomaly detection, they introduced the MVTec AD dataset, comprising 5354 high-resolution color images encompassing various object and texture categories. The dataset comprises both normal images, intended for training, and images with anomalies, designed for testing. These anomalies manifest in over 70 distinct types of defects, including scratches, dents, contaminations, and structural alterations. The authors also provided pixel-precise ground truth annotations for all anomalies.

The ability to recognize novel or anomalous images is a strength of human perception. However, machine learning systems often struggle with such tasks. There is a significant need for unsupervised algorithms capable of detecting anomalous regions, particularly in fields like manufacturing where supervised training samples are limited or the nature of defects is uncertain.

In this context, the authors identified the importance of datasets for anomaly detection and introduced the MVTec Anomaly Detection dataset to bridge this gap. They chose industrial inspection tasks as a suitable use case for the dataset, where defect-free images are employed for model training, and the model must detect anomalies during testing. The scarcity of defective samples in industrial processes makes unsupervised methods essential for accurate anomaly detection.

The MVTec Anomaly Detection dataset comprises 15 categories containing 3629 training images and 1725 testing images. Training images lack defects, while the testing set contains both defect-free images and images with various anomalies. Five categories cover different types of regular (carpet, grid) or random (leather, tile, wood) textures, while the remaining ten categories represent various types of objects. Some of these objects are rigid with a fixed appearance (bottle, metal_nut), while others are deformable (cable) or include natural variations (hazelnut). A subset of objects was acquired in a roughly aligned pose (e.g., toothbrush, capsule, and pill) while others were placed in front of the camera with a random rotation (e.g., metal_nut, screw, and hazelnut). The test images of anomalous samples contain a variety of defects, such as defects on the objects’ surface (e.g., scratches, dents), structural anomalies like distorted object parts, or defects that manifest themselves in the absence of certain object parts. In total, 73 different defect types are present, on average five per category.

All images were acquired using a 2048×2048 pixel high-resolution industrial RGB sensor in combination with two bilateral telecentric lenses with magnification factors of 1:5 and 1:1, respectively. Afterward, the images were cropped to a suitable output size. All image resolutions are in the range between 700×700 and 1024×1024 pixels. Each dataset image shows a unique physical sample. The authors did not augment images by taking multiple pictures of the same object in different poses. Since gray-scale images are also common in industrial inspection, three object categories (grid, screw, and zipper) are made available solely as single-channel images. The images were acquired under highly controlled illumination conditions. For some object classes, however, the illumination was altered intentionally to increase variability.

The pixel-precise ground truth labels were provided for each defective image region. In total, the dataset contains 1888 anomalous regions. All regions were carefully annotated and reviewed by the authors. During the acquisition of the dataset, the authors generated defects that were confined to local regions, which facilitated precise labeling of each anomaly. Additionally, pixels on the border of anomalies or lying in ambiguous regions were preferably labeled as anomalous. For locally deformed objects, annotations were created on the deformed area as well as in the region where the deformed object part is expected to be located. Some defects manifest themselves as missing parts. In these cases, the expected location of the part as anomalous was annotated.