Deep neural networks have become a standard tool in many modern perception tasks, including safety relevant applications like medical image processing and autonomous driving. The black box character of such nets however poses a number of challenges, as false positive and negative perceptions still occur frequently.
In this talk, we use intrinsic uncertainty measures for deep neural networks and use them for the prediction of misclassification. The basic concepts are thereafter applied to the semantic segmentation of street scenes. We also discuss the detection of false negatives based on decision rules and apply this to pedestrian detection.
Prof. Dr. Hanno Gottschalk, Bergische Universität Wuppertal