Alexander Timans

The ability of a model to quantify its uncertainty in predictions is a crucial component for its reliable deployment in real-life systems. This is particularly relevant for safety-critical applications such as autonomous driving, where a model's confidence in its predictions can be used to e.g. defer automated decision-making to a human user in case of high uncertainty, or be used for down-stream tasks such as anomaly detection. Computer vision is an important domain in the machine learning community which has high practical relevance for Bosch, such as in its mobility or smart automation projects. In this research project, we propose to investigate novel approaches to reliably and efficiently obtain uncertainty estimates for computer vision tasks such as object detection, object tracking, or different segmentation tasks. In particular, we are interested in investigating uncertainty methods which require minimal data assumptions and which are highly model-agnostic, and thus readily applicable for different vision tasks and model pipelines. Another desirable aspect is to obtain uncertainty estimates which contain some notion of theoretical guarantee on their validity or accuracy, such as a guaranteed coverage of ground-truth values or bounded risk. The connection to down-stream tasks such as anomaly detection, and to Bayesian and frequentist notions of error control are also of interest.