Reproduce “SeqGrowGraph: Learning Lane Topology as a Chain of Graph Expansions” on OpenResearch

About this paper

Mengwei Xie, Shuang Zeng, Xinyuan Chang, Xinran Liu, Zheng Pan, Mu Xu, Xing Wei

Researchers at Amap and Alibaba Group developed SeqGrowGraph, a transformer-based generative framework that models lane topology generation as a chain of graph expansions to robustly represent complex road structures like loops and bidirectional lanes. This approach achieved state-of-the-art F1-scores on nuScenes (L-F 56.4%, R-F 67.8%) and Argoverse 2 datasets, producing more stable and topologically accurate lane graphs than prior methods.

Key insights

• Representing graph generation as an incremental, human-inspired expansion process allows transformer-based sequence models to effectively learn and predict complex topological structures, including loops and bidirectional lanes.
• Assigning a higher weight to tokens representing point coordinates in the loss function improves overall performance, emphasizing the critical role of accurate node localization for correct topology and centerline definition.
• Depth-First Search (DFS) ordering for serializing subgraph descriptions proves most effective in capturing hierarchical relationships and local connectivity during the autoregressive generation process.

Summary by alphaXiv.