TopoGPT: Generative Lane Topology Reasoning with Geometry Prior

TopoGPT: Generative Lane Topology Reasoning with Geometry Prior

ECCV 2026

Jiahui Fu1 Zehao Huang Han Li1 Naiyan Wang Si Liu1,†

† corresponding author

1. Institute of Artificial Intelligence, Beihang University
Paper Code Weights Datasets

TopoGPT does not treat lane topology as a detect-then-match problem. Instead, it gives the model the ability to directly predict plausible road topology. We learn common road geometry from large-scale maps, then use this prior to predict geometrically consistent and structurally complete lane graphs from surround-view cameras.

Surround-view visual input for lane topology reasoning
Surround-view visual input
Previous lane detection paradigm visualization
Previous paradigm

Local lane instance detection

Independent lane detections
Pairwise topology matching
Misaligned endpoints and missing lanes
TopoGPT lane graph generation visualization
TopoGPT

Global lane graph generation

Learn road geometry from large-scale maps
Condition generation on camera BEV features
Complete, consistent lane graphs

The Model

TopoGPT uses an Autoregressive Lane Sequence Transformer to model lane graphs as token sequences. A two-stage training strategy first learns road-geometry regularities from map data, then transfers that prior to perception-conditioned lane graph prediction.

TopoGPT method visualization
Stage 1

Geometric Map-Prior Pre-training

Learn a geometry prior by autoregressively generating lane token sequences from large-scale map-derived scene tokens.

Stage 2

Perception-aware Alignment Fine-tuning

Align BEV perception tokens to the scene-token space so the pre-trained generator predicts realistic lane graphs from multi-view images.

The Data

Unlike image-map paired data, which requires strict spatial and temporal alignment, map-only data can be collected at much larger scale from offline maps. We gather 3.3M map scenes from multiple motion datasets and convert heterogeneous map annotations into a unified lane-token representation.

Dataset mix

Pre-training sources

Processing pipeline

Map to lane tokens

01 Scene Retrieval
02 Instance Merging
03 Geometric Sampling
04 Spatial Lexicographic Sort
05 Bézier-based Tokenizer
Output Lane Token Sequence

Lane topology quality

On OpenLane-V2, TopoGPT improves both lane-level matching and point-level graph quality. The four bar charts below report all lane metrics and graph metrics on subset A and subset B.

OpenLane-V2

Lane and graph metrics

Demo gallery

Qualitative lane graphs

Representative TopoGPT predictions visualize camera-conditioned lane graphs across diverse scenes.

TopoGPT qualitative lane graph prediction scene 1
TopoGPT qualitative lane graph prediction scene 2
TopoGPT qualitative lane graph prediction scene 3
TopoGPT qualitative lane graph prediction scene 4

Citation

Please cite our paper if TopoGPT is useful for your research.

@inproceedings{fu2026topogpt,
  title={Generative Lane Topology Reasoning via Autoregressive Model with Geometry Prior},
  author={Fu, Jiahui and Huang, Zehao and Li, Han and Wang, Naiyan and Liu, Si},
  booktitle={European Conference on Computer Vision},
  year={2026},
}