Out-of-Distribution detection in supervised image classification

Image classification has advanced significantly with deep learning in a well-controlled setting, where the test data are clean and from the same distribution as the training data. However, the deployment of deep learning models in the real world is still full of unknowns. More often than not, well-trained models can come across Out-of-Distribution (OoD) data that are sampled from a different distribution from that of the training data. For example, image classification models may encounter images that do not belong to any of the classes in the training data or simply In-Distribution (ID) images corrupted by noises and blurs. The problem of detecting Out-of-Distribution (OoD) data when using deep neural networks has been analyzed, and a simple yet effective way to improve the inter-class fairness and robustness of several popular OoD detection methods in the literature is proposed in this work. Most existing OoD detection methods do not take into account the misalignment among the output logit distributions given by multi-class classifiers. Through extensive experiments, it has been found that failure to address this issue not only can result in a biased OoD detector but also one that is vulnerable to label shift. The proposed class-wise detection algorithm not only mitigates significantly the bias but also maintains a similar OoD detection performance even in the presence of label shift in the test distribution. Moreover, few insights on how to extend the OoD detection analysis for the different problem of multi-label classification are suggested in the final part of the work.