Authors: Yifan Zhang and Yifei Zhang

Title: Synthetic Data Generation for Vision and LiDAR-Based Object Detection

Authors: Yifan Zhang and Yifei Zhang

Supervisors: Anders Heyden, Ali Nouri (Volvo Cars)

Examiner: Kalle Åström

Abstract:

This thesis addresses synthetic data generation for autonomous driving perception
systems, focusing on 3D Gaussian Splatting (3DGS) algorithms for novel view synthesis
with particular emphasis on accurate human modeling.We develop a comprehensive
framework utilizing the CARLA simulation environment to evaluate spatial
generalization capabilities across varying viewpoint distances. Our experiments
with state-of-the-art 3DGS variants reveal that all methods experience quality degradation
when rendering from novel perspectives, with performance boundaries becoming
evident at distances exceeding 3 meters from training viewpoints. The OmniRe
implementation demonstrates superior resilience, particularly for dynamic objects
such as pedestrians and vehicles.We introduce a modular trajectory generation
framework simulating realistic driving maneuvers and an object detection-based
evaluation methodology to quantify how rendering quality affects downstream perception
tasks. To overcome limitations in human representation from constrained
viewpoints, we develop an integrated pipeline combining text-guided Gaussian human
generation with scene graph modeling, enabling the creation of anatomically
accurate human models with consistent appearance from arbitrary viewpoints. Furthermore,
our system allows precise editing of vehicle and pedestrian positions,
as well as ego-vehicle trajectories, facilitating the deliberate design of hazardous
scenarios crucial for comprehensive autonomous driving testing. This methodology
significantly enhances synthetic data generation for safety-critical edge cases in
autonomous driving applications and provides insights for improving spatial generalization
in neural rendering techniques, ultimately contributing to more robust
perception systems for real-world driving scenarios.