Research map / Graph learning / Knowledge and temporal graphs
Topo2Seq: Enhanced Topology Reasoning via Topology Sequence Learning
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.
Topo2Seq trains a DETR-style lane segment decoder jointly with an auxiliary topology sequence decoder (randomized-order prompt-to-sequence learning) to improve lane topology reasoning from multi-view images, with the auxiliary decoder discarded at inference.
Problem
Lane topology reasoning from perspective/multi-view images with DETR-like lane graph predictors suffers from misaligned lane segment endpoints and weak long-range perception and connectivity modeling, while sequence-to-sequence alternatives are too slow at inference.
Approach
Use a sequence-to-sequence topology decoder only during training: key points predicted by the lane segment decoder are shuffled and fed as prompts to an autoregressive decoder that reconstructs an ordered DAG-based edge sequence, thereby supervising the lane segment decoder's BEV features and endpoint alignment without slowing inference.
Key results
- Topo2Seq with ResNet-50 for 48 epochs reaches 37.7 mAP and 29.9 TOP_lsls on OpenLane-V2 lane segment, versus LaneSegNet 36.4 mAP and 27.3 TOP_lsls at 48 epochs.
- On centerline perception at 48 epochs, Topo2Seq reaches 45.8 OLS* and 30.0 TOP_ll, above LaneSegNet's 43.3 and 27.3.
- Ablation shows randomized-order prompt-to-sequence learning raises mAP from 34.9 to 37.7 relative to the index 2 setting.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use Topo2Seq instead of LaneSegNet with 24 training epochs | OpenLane-V2 lane segment perception · OpenLane-V2 subset_A, 24 epochs | AP_ls: 31.9 -> 33.7 | Helped |
| use Topo2Seq instead of LaneSegNet with 48 training epochs | OpenLane-V2 lane segment perception · OpenLane-V2 subset_A, 48 epochs | TOP_lsls: 27.3 -> 29.9 | Helped |
| use Topo2Seq instead of LaneSegNet with 24 training epochs | OpenLane-V2 centerline perception · OpenLane-V2 subset_A, 24 epochs | OLS*: 40.7 -> 42.7 | Helped |
| use Topo2Seq instead of LaneSegNet with 48 training epochs | OpenLane-V2 centerline perception · OpenLane-V2 subset_A, 48 epochs | TOP_ll: 27.3 -> 30.0 | Helped |
| use randomized rather than ordered GT key point prompts (ablation) | Topo2Seq sequence learning · OpenLane-V2 subset_A lane segment ablation | TOP_lsls: 25.1 -> 27.8 | Helped |
| start prompt-to-sequence decoder interaction directly rather than first using randomized GT prompts (ablation) | Topo2Seq sequence learning · OpenLane-V2 subset_A lane segment ablation | mAP: 34.9 -> 33.8 | Hurt |
| add predicted key points and 24 epochs of decoder interaction after randomized GT prompts (ablation) | Topo2Seq sequence learning · OpenLane-V2 subset_A lane segment ablation | mAP: 34.9 -> 37.7 | Helped |
| add predicted key points and 24 epochs of decoder interaction after randomized GT prompts (ablation) | Topo2Seq sequence learning · OpenLane-V2 subset_A lane segment ablation | TOP_lsls: 27.8 -> 29.9 | Helped |
Surprising findings
- Interacting with the topology sequence decoder too early (using predicted key points before the lane decoder stabilizes) performed worse than using randomly ordered GT key point prompts.
Limitations
- Training uses a small total batch size of 4 on 4 A100 GPUs due to resource limits.
- Reported gains at 24 epochs are smaller than the 48-epoch results and rely on a two-stage schedule.
- Centerline and lane segment labels in OpenLane-V2 are misaligned, complicating comparisons.
Experimental setup
Datasets: OpenLane-V2 ImageNet Argoverse2 nuScenes
Benchmarks: OpenLane-V2 subset_A
Metrics: mAP AP_ls AP_ped TOP_lsls OLS* DET_l TOP_ll
Baselines: TopoNet MapTR MapTRv2 LaneSegNet VectorMapNet STSU Topo2D TopoMLP RoadPainter
Builds on: RoadNet LaneSegNet BEVFormer Pix2Seq DETR
Closest papers by meaning
- TopoLogic: An Interpretable Pipeline for Lane Topology Reasoning on Driving Scenes (NeurIPS 2024 · similarity 0.92)
- T2SG: Traffic Topology Scene Graph for Topology Reasoning in Autonomous Driving (CVPR 2025 · similarity 0.92)
- TopoMLP: A Simple yet Strong Pipeline for Driving Topology Reasoning (ICLR 2024 · similarity 0.92)
- TopoPoint: Enhance Topology Reasoning via Endpoint Detection in Autonomous Driving (NeurIPS 2025 · similarity 0.91)
- TopoCurve: Geometry-Aware Topology Reasoning via Bézier Curves in Autonomous Driving (NeurIPS 2026 · similarity 0.90)
- LaneSegNet: Map Learning with Lane Segment Perception for Autonomous Driving (ICLR 2024 · similarity 0.88)
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.