The MVTec Anomaly Detection Dataset: A Comprehensive Real-World Dataset for Unsupervised Anomaly Detection

Abstract The detection of anomalous structures in natural image data is utmost importance for numerous tasks the field computer vision. development methods unsupervised anomaly requires on which to train and evaluate new approaches ideas. We introduce MVTec dataset containing 5354 high-resolution color images different object texture categories. It contains normal, i.e., defect-free intended training with anomalies testing. manifest themselves form over 70 types defects such as scratches, dents, contaminations, various structural changes. In addition, we provide pixel-precise ground truth annotations all anomalies. conduct a thorough evaluation current state-of-the-art based deep architectures convolutional autoencoders, generative adversarial networks, feature descriptors using pretrained neural well classical vision methods. highlight advantages disadvantages multiple performance metrics threshold estimation techniques. This benchmark indicates that leverage networks outperform other deep-learning-based models show considerable room improvement.