Deep Evidential Uncertainty Estimation for Semantic Segmentation under Out-Of-Distribution Obstacles

In order to navigate safely and reliably in novel environments, robots must estimate perceptual uncertainty when confronted with out-of-distribution (OOD) obstacles not seen in training data. We present a method to accurately estimate pixel-wise uncertainty in semantic segmentation without requiring real or synthetic OOD examples at training time. From a shared per-pixel latent feature representation, a classification network predicts a categorical distribution over semantic labels, while a normalizing flow estimates the probability density of features under the training distribution. The label distribution and density estimates are combined in a Dirichlet-based evidential uncertainty framework that efficiently computes epistemic and aleatoric uncertainty in a single neural network forward pass. Our method is enabled by three key contributions. First, we simplify the problem of learning a transformation to the training data density by starting from a fitted Gaussian mixture model instead of the conventional standard normal distribution. Second, we learn a richer and more expressive latent pixel representation to aid OOD detection by training a decoder to reconstruct input image patches. Third, we perform theoretical analysis of the loss function used in the evidential uncertainty framework and propose a principled objective that more accurately balances training the classification and density estimation networks. We demonstrate the accuracy of our uncertainty estimation approach under long-tail OOD obstacle classes for semantic segmentation in both off-road and urban driving environments.