Long-Tailed Anomaly Detection with Learnable Class Names

We re-train all baselines on the proposed long-tailed configurations, using the github links of Cut & Paste, MKD, DRAEM, RegAD, UniAD and AnomalyGPT. By default, the training pipeline in the github link is followed. Both UniAD and LTAD use the architecture of EfficientNetB7 for fair comparison. For RegAD, please refer to Sec.6 for the training details. For AnomalyGPT, it trains the model on additional data. For example, when the downstream dataset is MVTec, additional VisA dataset is used for training. For fair comparison, we re-train AnomalyGPT on the proposed long-tailed configurations without using additional dataset.