Reproduce “AerialFusionMapNet: Online HD Map Construction with Aerial-Onboard BEV Fusion” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Daniel Lengerer, Mathias Pechinger, Klaus Bogenberger, Carsten Markgraf
Researchers from the Technical University of Applied Sciences Augsburg and Technical University of Munich developed AerialFusionMapNet, a framework that integrates aerial imagery with onboard sensor data for online HD map construction. A novel two-stage training strategy enables robust and accurate vectorized map element prediction, achieving a 5.9 mAP absolute improvement on the nuScenes geographic split and demonstrating stability against moderate aerial image misalignments.
Key insights
- The effectiveness of aerial-onboard fusion is more significantly influenced by the training strategy and representation alignment than by the raw capacity or complexity of the aerial encoder.
- Scenario-Consistent Rotation (SCR)-augmented aerial-only pretraining is crucial for creating a robust and geometrically consistent aerial reference representation, leading to substantial performance gains in subsequent multimodal alignment.
- Cross-View Supervision (CVS) effectively guides the onboard BEV representation to align with the structural priors from the aerial branch, enhancing feature coherence without modifying the inference architecture.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.