Efficient 3D object annotation via vision-derived pseudo-LiDAR and Vision Language Model (VLM) validation

To advance autonomous driving, accurate 3D object annotation is crucial for target recognition, environment perception, and high-precision map construction. However, producing high-quality 3D annotated data is costly and time-consuming. In particular, for sparse point cloud data, it is both labor-intensive and error-prone to annotate 3D objects. To address this challenge, this paper proposes an efficient automated annotation pipeline that integrates pseudo-point cloud generation with validation using a vision language model (VLM). Our approach supplements sparse point cloud data, generates pseudo-labels, and leverages a VLM model to validate and filter annotations, thereby creating a closed-loop automated system. Experiments on a real-world dataset collected by an autonomous vehicle demonstrate significant improvements in annotation accuracy and efficiency.