PseudoMapLabeler boosts HD map AI with +6.1 mAP via spatial clipping

Online HD map construction is critical for autonomous driving, but real-world deployment is hindered by the scarcity of labeled training data. To address this, researchers introduce PseudoMapLabeler, a teacher-student semi-supervised learning (SSL) framework that generates high-quality pseudo-labels from unlabeled data. The method first trains a teacher model on limited labeled data. It then leverages Beta-distribution-based confidence maps to assess the reliability of predicted map elements across temporal observations. Unlike conventional filtering that discards entire elements, PseudoMapLabeler applies a spatial clipping technique that selectively preserves high-confidence regions while removing unreliable segments. These refined map elements act as priors, improving the teacher model's predictions on unlabeled data in a second pass. The enhanced predictions become pseudo-labels to train a student model from scratch, followed by fine-tuning on the original labeled data.

Experiments on the nuScenes dataset show that the teacher-student framework with refined pseudo-labels improves performance by +6.1 mAP compared to training on labeled data alone in a low-label regime. This is a significant gain for HD map construction, where labeled data is expensive and difficult to obtain. The work also highlights the effectiveness of confidence-aware refinement over naive filtering methods, suggesting that spatial clipping can be a key component for robust SSL in vision-based mapping. Accepted at the ECCV 2026 DriveX Workshop on Foundation Models for Autonomous Driving, PseudoMapLabeler paves the way for more scalable and generalizable HD mapping systems, particularly in diverse environments where labeled data is scarce.