Talk at the SafeAI 2021

At the SafeAI workshop 2021, Maximilian Henne and Franziska Schwaiger will present their new paper »From Black-box to White-box: Examining Confidence Calibration under different Conditions«. Together with Karsten Roscher and Felippe Schmoeller da Roza from the Fraunhofer Institute for Cognitive Systems IKS and Fabian Küppers and Anselm Haselhoff from the Ruhr-West University of Applied Sciences, they develop and investigate methods for calibrating the confidence values that are assigned to the individual detections in 2D object detection.

For example, predictions with a confidence of 80 percent should be correct in eight out of ten cases on average. In particular, the researchers are looking at the influence of common post-processing methods of 2D object detection, such as non-maximum suppression, on the quality of the confidence calibration and are developing new methods to improve the calibration in terms of classification as well as localisation of the objects.