VectorMapNet: End-to-end Vectorized HD Map Learning

Autonomous driving systems require High-Definition (HD) semantic maps tonavigate around urban roads. Existing solutions approach the semantic mappingproblem by offline manual annotation, which suffers from serious scalabilityissues. Recent learning-based methods produce dense rasterized segmentationpredictions to construct maps. However, these predictions do not includeinstance information of individual map elements and require heuristicpost-processing to obtain vectorized maps. To tackle these challenges, weintroduce an end-to-end vectorized HD map learning pipeline, termedVectorMapNet. VectorMapNet takes onboard sensor observations and predicts asparse set of polylines in the bird's-eye view. This pipeline can explicitlymodel the spatial relation between map elements and generate vectorized mapsthat are friendly to downstream autonomous driving tasks. Extensive experimentsshow that VectorMapNet achieve strong map learning performance on both nuScenesand Argoverse2 dataset, surpassing previous state-of-the-art methods by 14.2mAP and 14.6mAP. Qualitatively, VectorMapNet is capable of generatingcomprehensive maps and capturing fine-grained details of road geometry. To thebest of our knowledge, VectorMapNet is the first work designed towardsend-to-end vectorized map learning from onboard observations. Our projectwebsite is available at\url{https://tsinghua-mars-lab.github.io/vectormapnet/}.