Preview

PDF, English Download (10MB) | Terms of use

Abstract

Depth estimation is a computer vision task that involves the estimation of distance to objects based on provided images and plays a crucial role in a wide variety of applications, such as autonomous driving. In many application areas, the decisions of Neural Networks (NNs) are safety-critical and incorrect ones can lead to serious consequences. However, traditional NNs do not provide any indication about the certainty of their predictions, which can lead to overconfidence, especially in situations that are not present in the training data. Bayesian Neural Networks (BNNs) extend the capabilities of traditional NNs by modeling uncertainty in their predictions. Since analytical Bayesian inference is often intractable for high-dimensional NNs, approximate approaches are used to realize BNNs. Monte Carlo Dropout (MCDO) and Deep Ensembles (DEs) are two BNN methods that have gained popularity due to their compatibility with established machine learning frameworks and straightforward application to NNs. Repulsive Last Layer Ensembles (RLLEs) present an alternative approach that offers the same compatibility and, additionally, focuses on resource efficiency, which is vital for complex, resource-intensive tasks such as depth estimation. To evaluate the different methods, we first perform an ablation study on a synthetic regression task and propose a new metric for epistemic uncertainty evaluation. MCDO has been found to demonstrate deficiencies in its capacity to predict epistemic uncertainties, while DEs and RLLEs both achieve good results. The application of the methods to depth estimation using established metrics yields similar results; yet they contradict the cross-validation based on artificially corrupted images. Overall, this work shows that ensemble-based BNN methods are well suited for regression tasks and even the resource-saving RLLEs perform well for the rather complex depth regression. This work is a step towards cost-efficient, generic NNs that have the ability to report uncertainty onto their predictions.