Core Concepts
SDMapとHDMapの事前情報を活用して性能を向上させるために、P-MapNetが提案されました。
Abstract
I. Introduction
Autonomous vehicles rely heavily on expensive HDMaps for navigation.
P-MapNet aims to improve online HDMap generation using SDMap and HDMap priors.
II. Methodology
A. SDMap Prior Module
Utilizes OpenStreetMap data to generate SDMaps.
Incorporates cross-attention module to enhance BEV features with SDMap prior.
B. HDMap Prior Module
Employs a masked autoencoder for refining initial HDMap predictions.
Achieves realistic far-seeing HDMaps in challenging scenarios.
III. Experiments
A. Dataset and Metrics
Evaluated on nuScenes and Argoverse2 datasets for different perception ranges.
Utilizes IoU, AP, and LPIPS metrics for evaluation.
B. Results
Outperforms existing methods in both short-range and long-range perception tasks.
Demonstrates the effectiveness of incorporating SDMap and HDMap priors.
IV. Ablation Study
1. Detailed Runtime Analysis
Shows the computational overhead of the HDMap prior module.
2. SD Map Prior Fusion Strategies
Cross attention approach demonstrates the most substantial gains in performance.
3. BEV-SdPrior Cross Attention Layers
Increasing transformer layers improves performance but may lead to overfitting.
V. Generalization Capability of HDmap MAE
Pre-training on different datasets shows generalization capability of the refinement module.
Stats
SDマップとHDマップの事前情報を活用することで、mIoUが最大18.73%向上しました。
HDマップ先行モジュールは計算量が多いですが、オプションです。