Deep Learning + Interpretable AI: Class imbalance and BEV feature focus in HD Map Perception

Forschungsthema/Bereich

Autonomous Driving, High Definition Maps, Deep Learning

Typ der Abschlussarbeit

Bachelor / Master

Startzeitpunkt

05.01.2026

Bewerbungsschluss

01.05.2026

Dauer der Arbeit

4-6 months

Beschreibung

Current state-of-the-art map construction methods such as MapTRv2 use sensor
data (360° surround view camera setup and LiDAR) to construct high definition
(HD) maps. These methods consists of two components the map encoder and
the map decoder. The map encoder extracts features from the sensor data and
transform them into a Bird’s Eye View (BEV) representation, whereas the map
decoder derives a map in polyline representation using transformer-based archi-
tectures from those BEV features. Visualizing the BEV features and predictions
shows that those models perform well on simple straight roads but struggle
with more complex scenarios such as intersections, which could be related to
class imbalance in the datasets. Also the focus of the Map decoder on the BEV
features is not researched yet and could give insights to possible class imbalance
problems.
The goal of this thesis is to develop and implement different methods to overcome
class imbalance in HD Map Perception and also use interpretable AI methods to
study the effect of those imbalances on the map perception task. The dataset
that will be used in this thesis is Argoverse 2.

The proposed thesis consists of the following parts:

+ Literature research about Class imbalance and Interpretable AI methods re-
garding HD Map Perception
+ Implementation of interpretable AI methods to visualize the attention of the
transformer architecture
+ Implementation of methods to overcome class imbalance within the HD map
elements

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

Link to the announcement:

• https://www.mrt.kit.edu/z/download/studip/Deep_Learning_Interpretable_AI_Class_imbalance_and_BEV_feature_focus_in_HD_Map_Perception.pdf

Voraussetzung

Voraussetzungen an Studierende

• Expericence with Neural Networks in Deep Learning context
• Experience with PyTorch, NumPy and Matplotlib
• Motivation and independent work style with the interest learning new things

Studiengangsbereiche

• Ingenieurwissenschaften
  Elektrotechnik & Informationstechnik
  Informatik
  Maschinenbau
  Mechatronik & Informationstechnik
  Mechanical Engineering
  Computer Science
  Mechatronics and Information Technology
• Naturwissenschaften und Technik
  Computational and Data Sience

Betreuung

Titel, Vorname, Name

Jonas, Merkert

Organisationseinheit

Institut für Mess- und Regelungstechnik (MRT)

E-Mail Adresse

jonas.merkert@kit.edu

Link zur eigenen Homepage/Personenseite

Website

Bewerbung per E-Mail

Bewerbungsunterlagen

• Lebenslauf
• Notenauszug
• Immatrikulationsbescheinigung

E-Mail Adresse für die Bewerbung
Senden Sie die oben genannten Bewerbungsunterlagen bitte per Mail an jonas.merkert@kit.edu


Zurück