A Unified Model for Multi-class Anomaly Detection - Supplementary Material - Zhiyuan You

We organize the supplementary material as follows. D conducts comprehensive ablation studies on the components of our approach. E presents more visualization results of the reconstructed features and qualitative results for Our task setting clearly differs from semantic AD. Unlike CIFAR-10, each category has normal and abnormal samples in MVTec-AD. We would like to model the joint distribution of normal samples across all categories .