Camilla Olshov and Armon Haghighi presenting their master thesis on Generation of Synthetic Image Degradation for Perception in Autonomous Driving

Examinator: Viktor Larsson
Advisors:  Alexandros Sopasakis and Abolfazl Chaman Motlagh

Abstract. 

Collecting data for autonomous driving can be both time-consuming and expensive, particularly when the data must include specific types of camera degradation. This thesis investigates whether generative AI, particularly GANs and diffusion models, can generate
synthetic camera degradations from existing clean images in a faster and more controlled manner. Different models and fine-tuning methods, specifically Textual Inversion and Low-Rank Adaptation (LoRA), were explored and compared. The experiments were primarily conducted using data provided by Qualcomm. However, because these data are confidential, the visual results presented in this thesis use the public NuScenes and WoodScape datasets, both commonly used in autonomous-driving research. To assess the realism of the generated degradations, an LPIPS-based score was developed and validated through a human perceptual survey. A range of methods was investigated, resulting in a final proposed method that generated realistic, localised degradations while largely preserving the background scene and demonstrated applicability across several camera types. Although output quality was not always consistent, the final method was computationally efficient and together with the perceptually supported scoring system, enabled the large-scale generation and automatic selection of convincing synthetic degradations. The combination of an efficient generation method and an effective evaluation metric establishes a practical foundation for future work.