PseudoMapLabeler: Confidence-Aware Pseudo-Label Generation for Semi-Supervised Online Mapping

A recent preprint on arXiv (2608.12600) presents PseudoMapLabeler, a semi-supervised learning framework that employs a teacher-student model to tackle the challenge of limited labeled training data in the development of online HD maps. Initially, a teacher model is trained using a small dataset with labels, and then Beta-distribution-based confidence maps evaluate the trustworthiness of predicted map components over time. Rather than discarding all unreliable elements, a spatial clipping method is utilized to retain high-confidence areas while eliminating less reliable parts. These enhanced map components are then used as priors to refine the teacher model's predictions for unlabeled data in a subsequent phase, ultimately aiming to improve generalization in various real-world scenarios. The paper can be accessed on arXiv with the identifier 2608.12600.