We will include here details on the NAGAD anomaly detection setting proposed in our paper "Red PANDA: Disambiguating Anomaly Detection by Removing Nuisance Factors"
The datasets we used to evaluate the NAGAD anomaly detection setting can be found in the following link: https://drive.google.com/drive/folders/13EaXyf4XQvOenHgOJL1sn9UMdWRTijZ9?usp=sharing
In each file, the key 'imgs' contain the samples; 'anom_label' describes the ground truth label (0 - a normal seen sample, 1 - an anomaly, 2 - normal "pseudo anomaly"). The other keys described the attributes directly. They are not part of the setting but are used to evaluate the used disentanglement approach.
We also supply in the script mvtec_augment_data.py which is used for the MVTec-AD based anomaly detection with nuisance factors benchmark (described in Sec.6 and App.A in the paper).