MapGCLR: Geospatial Contrastive Learning of Representations for Online Vectorized HD Map Construction
Researchers have developed MapGCLR, a novel self-supervised learning method to improve the construction of vectorized high-definition maps for autonomous vehicles. This approach enhances the representation of bird's-eye-view features by enforcing geospatial consistency between overlapping map segments using a contrastive loss function. By training on a larger unlabeled dataset with multi-traversal requirements, MapGCLR outperforms traditional supervised methods in downstream perception tasks and qualitative visualization. AI
IMPACT Enhances autonomous vehicle navigation and reduces costs associated with HD map creation.