Research map / Graph learning / Graph algorithms and clustering
LaneGraph2Seq: Lane Topology Extraction with Language Model via Vertex-Edge Encoding and Connectivity Enhancement
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Casts lane graph extraction from multi-camera images as sequence prediction, serializing vertices (DFS order) and edges into tokens for a Transformer decoder, with classifier-free guidance and nucleus sampling to improve connectivity.
Problem
Extracting lane graphs (centerline curves and their connectivity forming a DAG) from onboard sensors is challenging; prior two-stage DETR-style methods model edges weakly, and Graphormer-style Transformer graph methods only encode vertices explicitly.
Approach
Serialize a lane graph into a token sequence that jointly encodes vertices (DFS-ordered coordinates) and edges (child indices plus Bezier midpoints with split/EOE tokens), decode it auto-regressively from a BEV feature, and boost edge connectivity at inference via classifier-free guidance conditioned on the vertex sequence combined with nucleus sampling.
Key results
- On nuScenes with VoVNetV2 backbone (LaneGraph2Seq†), M-F 73.3, Detect 68.7, C-F 67.6, versus STSU's M-F 57.5 and C-F 56.0.
- With ResNet-50 on nuScenes, LaneGraph2Seq reaches M-F 64.1 and C-F 63.2, outperforming STSU and TPLR.
- On Argoverse 2 with VoVNetV2, LaneGraph2Seq† achieves M-F 68.0 and C-F 63.9.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use LaneGraph2Seq with VoVNetV2 instead of STSU | Lane graph extraction · nuScenes | M-F: 57.5 -> 73.3 | Helped |
| Use LaneGraph2Seq with VoVNetV2 instead of STSU | Lane graph extraction · nuScenes | Detect: 60.6 -> 68.7 | Helped |
| Use LaneGraph2Seq with VoVNetV2 instead of STSU | Lane graph extraction · nuScenes | C-F: 56.0 -> 67.6 | Helped |
| Use LaneGraph2Seq with VoVNetV2 instead of the ResNet-50 version | LaneGraph2Seq · nuScenes | M-F: 64.1 -> 73.3 | Helped |
| Use LaneGraph2Seq with VoVNetV2 instead of the ResNet-50 version | LaneGraph2Seq · Argoverse 2 | M-F: 61.7 -> 68.0 | Helped |
| Use BFS serialization instead of DFS (ablation) | LaneGraph2Seq · nuScenes | M-F: 73.3 -> 72.1 | Hurt |
| Use BFS serialization instead of DFS (ablation) | LaneGraph2Seq · nuScenes | Detect: 68.7 -> 67.2 | Hurt |
| Sort vertices by ascending xy coordinates instead of DFS (ablation) | LaneGraph2Seq · nuScenes | M-F: 73.3 -> 69.8 | Hurt |
| Sort vertices by ascending xy coordinates instead of DFS (ablation) | LaneGraph2Seq · nuScenes | Detect: 68.7 -> 61.7 | Hurt |
| Randomize vertex order instead of using DFS (ablation) | LaneGraph2Seq · nuScenes | M-F: 73.3 -> 70.2 | Hurt |
| Randomize vertex order instead of using DFS (ablation) | LaneGraph2Seq · nuScenes | Detect: 68.7 -> 61.3 | Hurt |
| Use 3 Transformer layers instead of 6 (ablation) | LaneGraph2Seq · nuScenes | M-F: 73.3 -> 70.8 | Hurt |
Surprising findings
- The model tends to terminate generation before predicting all edges, motivating the connectivity enhancement at inference.
Limitations
- Connectivity enhancement trades diversity against accuracy; alpha_c must be tuned (best at 4).
- Relies on BEV features from a LSS encoder pretrained on centerline segmentation.
Experimental setup
Datasets: nuScenes Argoverse 2
Benchmarks: nuScenes Argoverse 2
Metrics: M-P M-R M-F Detect C-P C-R C-F
Baselines: PINET Poly STSU TPLR TopoNet LaneGAP
Builds on: STSU LSS Graphormer classifier-free guidance
Closest papers by meaning
- Topo2Seq: Enhanced Topology Reasoning via Topology Sequence Learning (AAAI 2025 · similarity 0.88)
- Lane2Seq: Towards Unified Lane Detection via Sequence Generation (CVPR 2024 · similarity 0.88)
- LaneSegNet: Map Learning with Lane Segment Perception for Autonomous Driving (ICLR 2024 · similarity 0.88)
- Bezier Everywhere All at Once: Learning Drivable Lanes as Bezier Graphs (CVPR 2024 · similarity 0.87)
- TopoLogic: An Interpretable Pipeline for Lane Topology Reasoning on Driving Scenes (NeurIPS 2024 · similarity 0.87)
- T2SG: Traffic Topology Scene Graph for Topology Reasoning in Autonomous Driving (CVPR 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.