Master Thesis: Diffusion models for image-to-image translation

https://youtu.be/aUNiVxmiJDE


Background of thesis project


Verifying the software used for autonomous vehicles is a challenging problem. Part of the verification process is automated and cannot run on the vehicle in real time on the road due to security issues that have to be found during verification and testing.


Therefore, high-end simulation environments are crucial to test some of the modalities. The Carla open-source simulator is one of such environments. Since it is a simulation the visual input to the learning model differs significantly from the real-world environment. Thus, enhancing the quality of these input pictures is a promising direction researcher have been exploring with generative adversarial networks (GAN)[1,3]. Nevertheless, a reliable solution to consistently enhance and generate realistic images has yet to be realized with these systems.


The aim of this thesis is to use diffusion-based models [2] to create an image-to-image translation algorithm from images generated by the Carla simulator to real-world images.


We will first focus on understanding the theoretical background of diffusion models and how these methods can be applied to the image-to-image translation problem. Finally, we build an algorithm to solve the task.


If you find this proposal interesting send your application with CV and grades to:
leo.laine@volvo.com


Contact persons:
Leo Laine – Volvo GTT
tel: +46 31 323 53 11
mail: leo.laine@volvo.com


David Fitzek – Volvo GTT / Chalmers
mail: david.fitzek@volvo.com


Thesis Level: Master
Language: English
Starting date: 2nd January 2023
Number of students: 2


Tutor
Leo Laine – Volvo GTT
tel: +46 31 323 53 11
mail: leo.laine@volvo.com


David Fitzek – Volvo GTT / Chalmers
mail: david.fitzek@volvo.com