Map Learning

A successful deployment of automated vehicles in complex urban scenarios demands an excellent scene understanding and a reliable interpretation of the surroundings. To this end, neural networks can be trained with large amount of high-quality data. The manual annotation of this data is a very time-consuming and costly process, and the research community is seeking to (semi)-automate this process.

Towards this goal, we present Map Learning: A novel map-based data generation procedure for the training of neural networks. Our highly accurate localization in verified HD maps allows us to utilize semantic map information to generate a scalable amount of semantic ground truth data for autonomous driving applications without extra labeling effort. We leverage our multi-drive mapping algorithms to scale our data set and reach a sufficient data set size and variety to fulfil the task at hand. The data is generated by back-projecting 3D objects into the sensor measurements.


Frank Bieder, Haohao Hu, Johannes Schantz, Oguzahn Kirik, Florian Ries, Martin Haueis, Christoph Stiller. Map Learning: Ein Skalierbarer Ansatz zur Automatisierten Erstellung von Trainingsdaten unter Verwendung von HD Karten und Mehrfachbefahrungen. In 15. Workshop Fahrerassistenz Und Automatisiertes Fahren (FAS), Berkheim, Germany, October 2023. (Best Paper Award).