The path to a comprehensive examination of corner cases
An analysis of corner cases is one of the elementary topics in the project KI Data Tooling. The project goal of creating a comprehensive data kit for autonomous driving can only be achieved when these rarely occurring but often highly relevant scenarios are included and contemplated. Two recent publications illustrate the progress of the project with regard to this topic.
In a first step, an analysis of classifications and different abstraction levels of corner cases were examined in full detail. The results of those examinations are summarized in the paper „An Application-Driven Conceptualization of Corner Cases for Perception in Highly Automated Driving“ written by the University of Kassel, the Technical University of Braunschweig and the Forschungszentrum Informatik. The paper, which is accepted for publication on the Intelligent Vehicles Symposium (IV) 2021, is intended to contribute to a basic understanding of taxonomy and definitions in order to align further work on corner cases – within the project KI Data Tooling as well as beyond. Furthermore, the work provides an additional step forward: It extends the usual consideration of camera data by also analysing RADAR and LiDAR. The paper thus provides a better understanding of corner cases and is an important foundation for further work on the topic.
Continued work is described in the paper on „Out-of-distribution Detection and Generation using Soft Brownian Offset Sampling and Autoencoders“ by the University of Kassel accepted for publication on the CVPR Workshop “Safe Artificial Intelligence for Automated Driving”. Based on existing definitions of corner cases, it was examined to what extent these can also be generated synthetically. This is another important step to create a comprehensive database that must include critical and rare corner cases. Since these are rarely found in real data, it is essential to represent such scenarios synthetically. With the new "Soft Brownian Offset method" that was established in the project, corner cases (here, so-called out-of-distribution samples) are to be generated that lie at the edge or outside of the usual data distributions.
The aim of this work in KI Data Tooling is to produce a comprehensive catalogue of scenarios and their description, as well as methods and tools that can help with the automated detection and creation of corner cases. Both new publications illustrate the project's progress towards this goal.
Image: Universität Kassel