Bayesian Uncertainty Quantification with Synthetic Data

摘要

Image semantic segmentation systems based on deep learning are prone to making erroneous predictions for images affected by uncertainty influence factors such as occlusions or inclement weather. Bayesian deep learning applies the Bayesian framework to deep models and allows estimating so-called epistemic and aleatoric uncertainties as part of the prediction. Such estimates can indicate the likelihood of prediction errors due to the influence factors. However, because of lack of data, the effectiveness of Bayesian uncertainty estimation when segmenting images with varying levels of influence factors has not yet been systematically studied. In this paper, we propose using a synthetic dataset to address this gap. We conduct two sets of experiments to investigate the influence of distance, occlusion, clouds, rain, and puddles on the estimated uncertainty in the segmentation of road scenes. The experiments confirm the expected correlation between the influence factors, the estimated uncertainty, and accuracy. Contrary to expectation, we also find that the estimated aleatoric uncertainty from Bayesian deep models can be reduced with more training data. We hope that these findings will help improve methods for assuring machine-learning-based systems.