P-MapNet is an advanced online HDMap generation algorithm that enhances the performance of autonomous vehicles by incorporating both Standard Definition Map (SDMap) and High-Definition Map (HDMap) priors. The paper addresses the limitations of relying solely on HDMaps, which are expensive and not available in all regions. P-MapNet leverages SDMap priors to improve the initial map generation and HDMap priors to refine the output, making it more realistic and accurate, especially in challenging conditions such as occlusions and long-range sensing. Key contributions of P-MapNet include: 1. **SDMap Prior**: Extracts weakly aligned SDMap data from OpenStreetMap and encodes it as an additional conditioning branch, improving performance through attention-based architecture. 2. **HDMap Prior**: Uses a masked autoencoder to capture the prior distribution of HDMaps, refining the output to correct artifacts and improve perceptual metrics. 3. **Far-seeing Performance**: Significantly improves map generation performance over longer ranges, demonstrating effectiveness in real-world scenarios. The method is evaluated on the nuScenes and Argoverse2 datasets, showing improvements in both rasterized and vectorized output representations. P-MapNet can be switched between different inference modes to balance accuracy and efficiency, making it a versatile solution for autonomous driving applications. The code and models are publicly available for further research.P-MapNet is an advanced online HDMap generation algorithm that enhances the performance of autonomous vehicles by incorporating both Standard Definition Map (SDMap) and High-Definition Map (HDMap) priors. The paper addresses the limitations of relying solely on HDMaps, which are expensive and not available in all regions. P-MapNet leverages SDMap priors to improve the initial map generation and HDMap priors to refine the output, making it more realistic and accurate, especially in challenging conditions such as occlusions and long-range sensing. Key contributions of P-MapNet include: 1. **SDMap Prior**: Extracts weakly aligned SDMap data from OpenStreetMap and encodes it as an additional conditioning branch, improving performance through attention-based architecture. 2. **HDMap Prior**: Uses a masked autoencoder to capture the prior distribution of HDMaps, refining the output to correct artifacts and improve perceptual metrics. 3. **Far-seeing Performance**: Significantly improves map generation performance over longer ranges, demonstrating effectiveness in real-world scenarios. The method is evaluated on the nuScenes and Argoverse2 datasets, showing improvements in both rasterized and vectorized output representations. P-MapNet can be switched between different inference modes to balance accuracy and efficiency, making it a versatile solution for autonomous driving applications. The code and models are publicly available for further research.