Precise and Efficient Scene Understanding for Autonomous Driving with VLMs

Institut
Professur für autonome Fahrzeugsysteme
Typ
Masterarbeit /
Inhalt
experimentell / theoretisch /  
Beschreibung

The Problem:
Vision-Language Models (VLMs) are AI that are great at understanding a scene in a general way, like describing what's in a picture. However, for a self-driving car, this isn't enough. These models have two big weaknesses:

  1. They aren't very precise at pinpointing exactly where objects are.

  2. They are large and slow, making them unsuitable for the split-second decisions needed for driving.

 

The Goal:
We want to create a detection system for self-driving cars that is both smart (understands context like a VLM) and precise & fast (can quickly and accurately locate objects).

 

The Plan:
We will adapt existing VLMs in two key steps:

  1. Specialized Training: We will fine-tune a VLM using driving-specific data. This teaches it to be much better at the precise task of locating cars, pedestrians, and other critical objects on the road.

  2. Model Compression: We will then use a technique called "knowledge distillation" to transfer the understanding from the large, slow VLM into a much smaller and faster model. Think of it as training a compact, efficient student model with the knowledge of a large, smart teacher.

 

The Result:
The final product will be a lightweight, real-time object detector that doesn't just see objects, but understands the scene with the intelligence of a VLM, all while being fast and accurate enough for safe autonomous driving.

 

Key Facts

Type: MA, also for Informatics students
Starting Date: Immediately
Supervisor: Prof. Dr.-Ing. Johannes Betz   
Advisor: Yuchen Zhang, M.Sc   
Programming Language: Python   
Language: English
Required Knowledge: Python + Computer Vision/Object Detection

 

Work can begin immediately. If you are interested, simply send an email with your CV and academic transcript to yuchen2.zhangtum.de ;)

Tags
AVS Zhang
Möglicher Beginn
sofort
Kontakt
Yuchen Zhang
yuchen2.zhangtum.de