Research overview

Structure-preserving scene generation for autonomous-driving data augmentation

A recruiter-facing overview of the bachelor thesis project: generate rare driving-scene variations while preserving road geometry and objects.

Overview diagram
Goal

Generate night, rain, or other rare driving conditions while keeping road layout, objects, depth, and edges useful for driving-related tasks.

Role

Bachelor thesis; model design, implementation, evaluation pipeline, documentation, and model-weight release.

Evidence

Depth RMSE improved from 33.02 to 27.77; Object Preservation F1 improved from 0.0889 to 0.1071.

Tech

Diffusion models, ControlNet, Stable Diffusion, CLIP/open_clip, Qwen3-VL, Metric3D, Waymo, NVIDIA RTX 5090.

Code Model weights Detailed project siteBack to portfolio