Reproduce “SparScene: Efficient Traffic Scene Representation via Sparse Graph Learning for Large-Scale Trajectory Generation” 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
Xiaoyu Mo, Jintian Ge, Zifan Wang, Chen Lv, Karl Henrik Johansson
SparScene introduces an efficient traffic scene representation method that leverages HD-map lane topology to construct sparse, structure-aware interaction graphs for large-scale multi-agent trajectory generation. The framework achieves a 23-fold improvement in inference speed (5 ms for 32 agents) and 96% parameter reduction compared to state-of-the-art models on WOMD, while maintaining competitive prediction accuracy (minFDE of 1.1834m) and scaling to thousands of agents simultaneously.
Key insights
- Meaningful interactions in large-scale traffic scenes are inherently sparse and strongly structured by road topology, rather than solely by geometric proximity.
- Explicitly guiding interaction graph construction using lane connectivity from HD-maps effectively reduces computational complexity without sacrificing critical behavioral information.
- A hierarchical message passing strategy, incorporating local traffic dynamics into lanes (TiL), propagating structured lane context to agents (L2A), and modeling agent-to-agent interactions (A2A) based on lane reachability, optimizes performance and efficiency.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.