MS-AD: a Metal Stamping Dataset for Visual Anomaly Detection under Real-World Conditions

Visual anomaly detection (VAD) has achieved strong results in automated quality control, yet most benchmarks are acquired under controlled laboratory conditions with static backgrounds and uniform illumination. This limits their ability to represent the visual complexity of active production lines. We introduce the Metal Stamping Anomaly Detection (MS-AD) dataset, collected directly from a high-speed metal stamping process under real operational conditions. MS-AD captures dynamic backgrounds, illumination changes caused by reflective surfaces, between training and test data. The dataset contains 1,480 images of a metal strip component, with four pixel-wise annotated defect categories. We benchmark representative state-of-the-art anomaly detection methods and show that performance decreases under operational variability, highlighting the need for realistic evaluation settings for industrial visual inspection.