VectorMapNet: End-to-end Vectorized HD Map Learning

Autonomous driving relies heavily on high-definition (HD) semantic maps to navigate complex urban environments. Traditional methods depend on manual annotation, which is labor-intensive and lacks scalability, limiting deployment in real-world scenarios. Recent deep learning approaches have attempted to generate maps via dense rasterized segmentation, but these often lack instance-level information and require heuristic post-processing, leading to inaccuracies and inefficiencies.

In response, Yicheng Liu and colleagues introduce VectorMapNet, a novel end-to-end framework that directly predicts sparse, vectorized map elements—specifically, polylines—using onboard sensors such as cameras and LiDAR. The core idea is to convert multi-modal sensor data into bird’s-eye view features, then detect map elements via deformable detection transformers, and finally generate detailed polylines through autoregressive sequence models. This approach eliminates the need for pixel-based segmentation and heuristic post-processing, streamlining the entire mapping pipeline.

Extensive experiments on nuScenes and Argoverse2 datasets demonstrate that VectorMapNet outperforms previous state-of-the-art methods by at least 14.2 mAP. Qualitative results show the model's ability to capture fine geometric details, such as sharp corners and complex road boundaries, producing comprehensive and continuous maps suitable for downstream tasks like motion forecasting and path planning. These advances significantly enhance the fidelity, scalability, and real-time applicability of HD maps in autonomous driving systems.

The significance of this work lies in its potential to transform map learning from a manual, pixel-based process into an automated, vectorized, end-to-end pipeline. This not only reduces labor costs but also improves map accuracy and utility, enabling safer and more reliable autonomous vehicles. The technical innovations—multi-modal fusion, deformable detection, and autoregressive polyline generation—set new benchmarks in the field.

Looking ahead, future research will focus on integrating temporal data for dynamic mapping, reducing reliance on large annotated datasets, and optimizing computational efficiency for real-time deployment. Overall, VectorMapNet marks a pivotal step toward fully autonomous, scalable, and precise map construction, promising broad impacts across industry and academia.