Unsupervised Domain Adaptation (UDA) for 3D object detection in autonomous driving faces challenges due to various sources of domain shift, such as differences in LiDAR resolution, the use of synthetic versus real-world data, scenery variations, and sensor configurations (e.g., sensor placement and number). This thesis systematically investigates these domain shifts through controlled experiments using synthetic datasets generated via the CARLA simulator, enabling precise isolation and quantification of each factor. To facilitate these experiments, two software tools are introduced: carlaSceneCollector, designed for efficient synthetic data generation, and rosbag2nuScenes, which converts ROSBag data into the widely adopted nuScenes format. The study emphasizes two critical sources of domain shift: LiDAR resolution and the synthetic-to-real data shift. It identifies saturation effects at intermediate LiDAR resolutions (32–64 channels) and analyzes how varying resolution shifts impact detection performance, particularly noting the disproportionate effects on smaller objects. It evaluates various performance metrics, highlighting the robustness of the NuScenes Detection Score (NDS) compared to traditional metrics like mean Average Precision (mAP). Simultaneously, the synthetic-to-real domain shift is analyzed through systematic comparisons across the nuScenes, adaScenes, and carlaScenes datasets. This reveals that synthetic-to-real differences significantly surpass the impact of LiDAR resolution shifts, underscoring profound discrepancies between simulated and real-world LiDAR point clouds. The thesis further addresses limitations in the default voxelization settings of the CenterPoint model by proposing adaptive voxelization techniques and structural enhancements, enhancing model adaptability across resolutions. Finally, it examines real-world datasets like nuScenes, highlighting their complexity and diversity as key factors in achieving robust model performance and improved generalization.
Read
