École Nationale Supérieure de
Techniques Avancées

Abstract

Standard deep learning methods cannot provide information about the reliability of their predictions that is crucial for decision making and optimization in many fields of science and engineering. To alleviate this issue, including Bayesian Neural Networks (BNNs) and ensembling NNs are invented. Semantic Understanding of complex urban street scenes has been a popular topic, mainly applied in autonomous driving. However, no current dataset takes into account the uncertainty estimation of semantic segmentation yet. In this paper, we present ongoing work on a new large-scale virtual urban scenes dataset (City Anyverse) to assess the performance of the uncertainty quantification (UQ) algorithm and propose a UQ algorithm for different vision tasks. City Anyverse comprises a large, diverse set of virtual urban street scenes with different levels of weather conditions. All images have high-quality pixel-level annotations, including 3500 images in the train set, 500 in the validation set, and 4500 in the test set with weather conditions and OOD samples. The novel UQ algorithm based on one versus all is the extended work of OVNNI, which provides an idea for UQ. Keywords: Uncertainty quantification, Urban street scenes dataset, Semantic segmentation, OOD detection