Transformer-Based 3D Object Detection: Leveraging Attention for Uncertainty Estimation

Institut

Lehrstuhl für Fahrzeugtechnik

Typ

Masterarbeit /

Inhalt

experimentell / theoretisch /  

Beschreibung

Autonomous vehicles rely on object detectors to perceive their environment. Recently, transformer-based 3D detectors have emerged as the most performant class of architectures. While achieving a high mean Average Precision (mAP) is valuable for autonomous driving, another crucial property is uncertainty awareness. An uncertainty-aware detector not only predicts the maximum a posteriori estimate, but also provides a measure of confidence for each prediction. Reliable uncertainty estimates can significantly improve downstream tasks such as object tracking and motion prediction.

We already have uncertainty-aware object detectors available for both epistemic and aleatoric uncertainty and use aleatoric uncertainty estimation in our research vehicle. The goal of this thesis is to apply existing uncertainty estimation techniques to transformer-based architectures and to explore how the attention mechanism can be specifically leveraged to enhance uncertainty estimation. This serves the mid-term goal of deploying a more performant and modern object detector on our research vehicle EDGAR; however, this is outside the scope of this thesis.

What you’ll do:

• Review and analyze the leading transformer-based 3D object detectors
• Implement and evaluate epistemic and aleatoric uncertainty estimation. You will transfer the uncertainty estimation techniques from our existing networks to a selected transformer architecture.
• Explore possibilities to exploit the transformer architectures for better calibrated uncertainty estimates

Voraussetzungen

• Curiosity for autonomous driving and machine learning
• Experience with Python and Pytorch will definitely be helpful
• An engaged and independent working attitude

Tags

FTM Studienarbeit, FTM AV, FTM Schroeder, FTM Informatik

Möglicher Beginn

sofort

Kontakt

Cornelius Schröder, M.Sc.
cornelius.schroedertum.de

Ausschreibung