Open set recognition through Monte Carlo dropout-based uncertainty

Open set recognition has received much attention in recent years. In this paper, we present a novel open set recognition method that is able to obtain improved recognition by applying Monte Carlo Dropout to capture uncertainty in order to yield high quality predicted probabilities. Experimental results on six benchmark datasets show that our method gives better open set recognition performance than other stateof-the-art methods, with at least 6.4%, 3.9%, 2.9% and 1.0% performance increase in AUROC on the challenging object datasets CIFAR-10, CIFAR+10, CIFAR+50 and TinyImageNet respectively. We also perform an analysis on the benefits of combining predictive uncertainty with an EVT-based open set recognition model which indicates that Monte Carlo Dropout-based uncertainty allows to obtain high quality predicted probabilities and to learn more accurate open set recognition scores. This, in turn, helps to reduce the overlap between known and unknown classes, thus making them more separable.