Ing. Lukáš Neumann, Ph.D.

Accurate computer vision algorithms are a critical component of safe and robust autonomous driving. One of the limitations that currently hinders exploitation of modern computer vision algorithms in real-world scenarios is the scarcity of labelled training data, owing to the fact that human labelling is very time-consuming and therefore costly, which in turn causes potential safety concerns for autonomous driving because as a result many rare (aka long-tail) events are not covered in the training. The goal of this project is to train modern computer vision algorithms, such as 3D object detectors, through weakly supervised learning without relying on any human annotations in the process. This will allow the algorithms to be trained and constantly improved using unlabelled data which are readily available in quantities orders of magnitude higher than data with human annotations.