P-MapNet: Far-seeing Map Generator Enhanced by both SDMap and HDMap Priors

• DOI: 10.48550/arxiv.2403.10521
• Publication Date: 2024-03-15
• Authors: Zhou Jiang; Zhenxin Zhu; Pengfei Li; Huan-ang Gao; Tianyuan Yuan; Yongliang Shi; Hang Zhao; Hao Zhao
• Keywords: FOS: Computer and information sciences; Computer Vision and Pattern Recognition (cs.CV); Computer Science - Computer Vision and Pattern Recognition

Autonomous vehicles are gradually entering city roads today, with the help of high-definition maps (HDMaps). However, reliance on HDMaps prevents autonomous from stepping into regions without this expensive digital infrastructure. This fact drives many researchers to study online HDMap generation algorithms, but performance these algorithms at far is still unsatisfying. We present P-MapNet, in which letter P highlights that we focus incorporating map priors improve model performance. Specifically, exploit both SDMap and HDMap. On one hand, extract weakly aligned OpenStreetMap, encode it as an additional conditioning branch. Despite misalignment challenge, our attention-based architecture adaptively attends relevant skeletons significantly improves other a masked autoencoder capture prior distribution HDMap, can serve refinement module mitigate occlusions artifacts. benchmark nuScenes Argoverse2 datasets. Through comprehensive experiments, show that: (1) performance, using rasterized (by up $+18.73$ $\rm mIoU$) vectorized $+8.50$ mAP$) output representations. (2) perceptual metrics by $6.34\%$. (3) P-MapNet be switched different inference modes covers accuracy-efficiency trade-off landscape. (4) far-seeing solution brings larger improvements longer ranges. Codes models publicly available https://jike5.github.io/P-MapNet.