Learning from Maps: Scalable Ground Truth Generation in Autonomous Driving

Reliable map perception is essential for scalable autonomous driving, but modern perception systems require large amounts of costly manual annotations. This work presents a scalable approach for automatic training data generation using HD maps as supervision across multiple sensor modalities and tasks. The proposed methods enable online HD map construction, perspective map perception, and cross-modal learning with camera and LiDAR data while significantly reducing manual labeling effort.