SOR: Semi-Automatic Object Box and Road Element Annotation Pipeline

Accurate 3D bounding boxes and road element labels are critical for autonomous driving systems, but creating high-quality annotations faces significant challenges due to the domain gap problem in point cloud data and inefficient annotation workflows. This paper presents SOR (Semi-automatic Object box and Road element annotation), a novel interactive annotation pipeline that effectively addresses these challenges. SOR introduces three key innovations: (1) an interactive semi-automatic object detection algorithm achieving cross-domain adaptation through minimal human interaction; (2) a high-fidelity conversion mechanism from 3D point cloud scenes to 2D bird's-eye view (BEV) that reduces annotation complexity; and (3) improved pavement element labeling algorithm based on MapTRv2. Extensive experiments on the Waymo Open Dataset and nuScenes dataset demonstrate SOR's superior performance. Our Refine Model achieves theoretical optimality (100% AP) on the Waymo dataset while reducing Average Center Translation Error and Average Rotation Error by 50% and 56% respectively compared to state-of-the-art methods. For road element annotation, SOR achieves 74.8% mean AP on the nuScenes dataset, outperforming existing methods by up to 25.9%. These results confirm that SOR substantially reduces annotation costs and provides an efficient solution for large-scale point cloud annotation in autonomous driving applications, potentially accelerating the development and deployment of autonomous driving technology.