Inference Over Distribution of Posterior Class Probabilities for Reliable Bayesian Classification and Object-Level Perception

State of the art Bayesian classification approaches typically maintain a posterior distribution over possible classes given available sensor observations (images). Yet, while these approaches fuse all classifier outputs thus far, they do not provide any indication regarding how reliable the posterior classification is, thus limiting its functionality in terms of autonomous systems and robotics. On the other hand, current deep learning based classifiers provide an uncertainty measure, thereby quantifying model uncertainty. However, they do so on a single frame basis and do not consider a sequential framework. In this letter, we develop a novel approach that infers a distribution over posterior class probabilities, while accounting for model uncertainty. This distribution enables reasoning about uncertainty in the posterior classification and, therefore, is of prime importance for robust classification, object-level perception in uncertain and ambiguous scenarios, and for safe autonomy in general. The distribution of the posterior class probability has no known analytical solution; thus, we propose to approximate this distribution via sampling. We evaluate our approach in simulation and using real images fed into a convolutional neural network classifier.