Out-of-Distribution Detection without Class Labels

Anomaly detection methods identify samples that deviate from the normal behavior of the dataset. It is typically tackled either for training sets containing normal data from multiple labeled classes or a single unlabeled class. Current methods struggle when faced with training data consisting of multiple classes but no labels. In this work, we first discover that classifiers learned by self-supervised image clustering methods provide a strong baseline for anomaly detection on unlabeled multi-class datasets. Perhaps surprisingly, we find that initializing clustering methods with pre-trained features does not improve over their self-supervised counterparts. This is due to the phenomenon of catastrophic forgetting. Instead, we suggest a two stage approach. We first cluster images using self-supervised methods and obtain a cluster label for every image. We use the cluster labels as "pseudo supervision" for out-of-distribution (OOD) methods. Specifically, we finetune pretrained features on the task of classifying images by their cluster labels. We provide extensive analyses of our method and demonstrate the necessity of our two-stage approach. We evaluate it against the state-of-the-art self-supervised and pretrained methods and demonstrate superior performance.