Reproduce “Graph-based Topology Reasoning for Driving Scenes” 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

Tianyu Li, Li Chen, Huijie Wang, Yang Li, Jiazhi Yang, Xiangwei Geng, Shengyin Jiang, Yuting Wang and 5 more

TopoNet jointly predicts lane centerlines, lane connections, and links between lanes and traffic elements from vehicle-camera images. Its graph-based feature interaction lets lane representations receive type-specific messages from connected lanes and traffic-element queries. On OpenLane-V2 subset_A, it reports 35.6 OLS versus 26.0 for MapTR*, under a ResNet-50, 24-epoch comparison; this measures scene-structure prediction, not planning.

Key insights

  • The Scene Graph Neural Network (SGNN) effectively integrates heterogeneous features from 2D perspective view traffic elements and 3D BEV lanes into a unified feature space, enabling comprehensive topological reasoning.
  • Incorporating explicit semantic prior knowledge through the Scene Knowledge Graph (SKG), using class-specific weights for traffic elements and connection-order-specific weights for lanes, significantly enhances message passing and relationship prediction.
  • Jointly modeling both lane-to-lane connectivity and lane-to-traffic-element assignments within a single, end-to-end framework is crucial for a coherent and robust understanding of the complex 'road genome'.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.