Back to the Basics: Revisiting Out-of-Distribution Detection Baselines
This work addresses the problem of reliable OOD detection for image classification systems, but it is incremental as it revisits and refines basic methods rather than introducing new paradigms.
The paper tackles out-of-distribution (OOD) image detection by evaluating simple methods using only predictions or representations from pre-trained classifiers like ResNet-50 and Swin Transformer, finding that methods incorporating learned representations outperform those based solely on predictions. It advocates for a simple K-nearest neighbors distance approach in representation space, which has been neglected in prior studies.
We study simple methods for out-of-distribution (OOD) image detection that are compatible with any already trained classifier, relying on only its predictions or learned representations. Evaluating the OOD detection performance of various methods when utilized with ResNet-50 and Swin Transformer models, we find methods that solely consider the model's predictions can be easily outperformed by also considering the learned representations. Based on our analysis, we advocate for a dead-simple approach that has been neglected in other studies: simply flag as OOD images whose average distance to their K nearest neighbors is large (in the representation space of an image classifier trained on the in-distribution data).