SeqGrowGraph: Chain-of-Graph Expansion for Precise Lane Topology

ICCV (International Conference on Computer Vision) is a top‑tier computer‑vision conference. This year the conference will be held in Hawaii, USA, with an acceptance rate of 24% (2698/11239 submissions). The Gaode Vision Center team contributed five accepted papers.

Introduction

Accurate lane topology is a core element for high‑precision autonomous‑driving systems because it directly determines a vehicle’s understanding of the road environment and the correctness of path planning. Traditional methods struggle with complex non‑linear topologies such as roundabouts, multi‑direction lanes, and bidirectional single lanes, limiting their performance in real‑world scenarios.

To address these challenges, the Gaode Vision Center proposes SeqGrowGraph , a new framework that generates lane topology by a chain‑wise graph expansion process, inspired by how humans draw maps: starting locally and gradually extending to the whole map.

Paper title: SeqGrowGraph: Learning Lane Topology as a Chain of Graph Expansions

Paper link: (will be updated in the comments)

GitHub: (will be updated in the comments)

Approach

Sequence Construction

SeqGrowGraph models lane graphs as directed graphs G=(V,E), where V are intersections or key topology nodes and E are lane centerlines. The graph is built incrementally: each step introduces a new node with its spatial position and updates the adjacency matrix A from n×n to (n+1)×(n+1), encoding connectivity (upper‑triangular part for “from” edges, lower‑triangular for “to” edges). A geometric matrix M, represented by quadratic Bézier curves, updates the shape of the centerlines.

A Transformer‑based autoregressive model predicts these expansions in a depth‑first‑search order. Unlike DAG‑based methods, SeqGrowGraph can flexibly model complex real‑world road structures—including loops, bidirectional lanes, and non‑trivial topologies—without extensive pre‑ or post‑processing.

The process is illustrated below:

Each expansion step adds a new node (intersection) and its connected centerlines, updating the adjacency matrix rows and columns accordingly. The full lane graph is obtained by sequentially merging these sub‑graphs.

Model Structure

The method first uses a BEV encoder (LSS) to project features extracted from surround‑view images onto the bird‑eye‑view plane. After converting the lane graph to a sequence, a Transformer decoder generates the sequence token‑by‑token. The overall architecture is shown below:

Experiments

We compare SeqGrowGraph with state‑of‑the‑art methods on the nuScenes dataset using both the default split and the PON split. SeqGrowGraph achieves higher F1 scores on landmarks and reachability, and maintains a performance lead even when there is no scene overlap between training and test sets.

Qualitative visualizations on nuScenes show that SeqGrowGraph produces stable and robust lane graphs, effectively merging redundant points from real‑world data and outperforming competing methods.

Conclusion

SeqGrowGraph is an innovative autoregressive framework for lane‑graph modeling. By incrementally growing the graph as a sequence, it captures global structure while handling complex topologies such as loops and bidirectional lanes that challenge traditional DAG‑based approaches. Extensive experiments on nuScenes and Argoverse 2 demonstrate superior topological accuracy and network completeness, advancing lane‑graph construction for reliable autonomous driving.