ResAD: A Simple Framework for Class Generalizable Anomaly Detection

Because normal feature representations vary significantly across classes, this will cause the widely studied one-for-one AD models to be poorly class-generalizable ( i.e., performance drops dramatically when used for new classes).