Self-driving cars are the next step towards safe and convenient travel, but, as with all machine learning applications, require loads of training data. It would be desireable if the Freie Universität Berlin could us readily available datasets to prototype new machine learning models instead of creating their own dataset from test drives of one of their self-driving-cars.

Multiple traffic light detection models were trained on the popular datasets BSTLD and DTLD using the Tensorflow Research repository. The evaluation revealed that predictive power achieved in one dataset generally transfers over to another similar dataset with minimal loss in performance. Neither geographical differences between datasets (e.g. traffic lights at the beginning or the end of an intersection) nor architecture choices seem to impact this result. Some ideas to further reduce performance penalties are given for future work.