Enhancing deep learning based classifiers using out of distribution data detection

Abstract

In this work, three out-of distribution detection methods are implemented, evaluated and compared on several common benchmarks (different natural image datasets), as well as on the ImageNet-O dataset, a novel dataset that has been created to aid research in OOD detection for ImageNet models. In this thesis, we also investigate the effect of label space size on the OOD detection performance, for that we used three different in-distribution datasets (CIFAR-10, CIFAR-100 and ImageNet-1K), and we showed that the performance degrades rapidly as the number of in-distribution classes increases. We concluded by proposing a method that surpasses the three previous methods in detection performances and by creating a web user interface to test out our OOD detection method.