LIE: LiDAR-only HD Map Construction via Knowledge Distillation
A new method called LIE has been introduced by researchers for constructing high-definition maps in autonomous driving, relying solely on LiDAR technology. In contrast to camera-based techniques that do not offer depth information, LIE employs knowledge distillation from a teacher branch, integrating LiDAR features with 2D intensity maps for enhanced supervision. This approach surpasses the leading camera models on the nuScenes dataset by 8.2% in mean Intersection over Union (mIoU) and demonstrates resilience over extended distances as well as in difficult weather and lighting scenarios.
Key facts
- LIE is a LiDAR-only semantic map construction method
- Uses knowledge distillation to handle lack of dense semantic cues
- Teacher branch fuses student LiDAR features and 2D intensity map
- Outperforms all single-modality approaches on nuScenes
- 8.2% higher mIoU than state-of-the-art camera-based model
- Robust over long ranges and under challenging weather and lighting
- Published on arXiv with ID 2605.01478
- Method is efficient
Entities
—