Development of a Map Deprecation Framework for Vectorized HD Maps (Software Engineering)

Job vacancy: From now on

Description

Modern HD map construction methods, such as MapTRv2, use sensor data from
360° surround-view cameras and LiDAR to generate detailed, high-definition
vector maps. These methods extract features from sensor data, transform them
into a Bird’s Eye View (BEV) representation, and produce maps in a polyline format
using transformer-based architectures. However, the accuracy of these maps
depends heavily on precise labels. In practice, label quality can be compromised
due to road construction, outdated data, or human labeling errors.
HD map change detection aims to identify misalignments between map data and
recent sensor data when maps are outdated or inaccurate. While neural networks
can detect these issues, they require large amounts of training data. Currently,
the Trust but Verify (TbV) dataset is the only publicly available benchmark for
this task. Because real-world map changes are relatively rare, collecting enough
data is challenging. One solution is to artificially introduce map changes — for
example, simulating construction sites, lane closures, or changes in lane topology
— and use these modified maps to train detection models.

Your Role
In parallel with an ongoing master’s thesis on the impact of map label noise on
map perception, we are looking for a HiWi student assistant to:
• Extend the set of artificial map errors currently being developed
• Design and implement a Map Deprecation Framework in Python
• Integrate support for the Argoverse2, Trust but Verify, and NuScenes datasets
• Prepare the framework for open-source release on GitHub

I am happy to answer any questions you might have. Feel free to ask for an
appointment or directly ask at my office!