Reproduce “VectorNet: Encoding HD Maps and Agent Dynamics from Vectorized Representation” 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
Jiyang Gao, Chen Sun, Hang Zhao, Yi Shen, Dragomir Anguelov, Congcong Li, Cordelia Schmid
Waymo LLC and Google Research's VectorNet processes vectorized representations of High Definition maps and agent trajectories using hierarchical Graph Neural Networks for behavior prediction, achieving a 12% improvement in displacement error at 3 seconds on the Argoverse dataset while reducing computational FLOPs by 99.6% compared to rasterization-based methods.
Key insights
- Directly processing vectorized data preserves continuous geometric information and the inherent structure of road networks and trajectories, leading to better prediction accuracy.
- A hierarchical Graph Neural Network (GNN) architecture effectively captures both local polyline features and global interactions between different scene elements.
- Self-supervised learning through masked feature reconstruction significantly improves the model's ability to learn robust and generalizable contextual representations.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.