nearest neighbors is a successful and long-standing technique for anomaly
detection. Significant progress has been recently achieved by self-supervised
deep methods (e.g. RotNet). Self-supervised features however typically
under-perform Imagenet pre-trained features. In this work, we i