Research map / Graph learning / Knowledge and temporal graphs
T2SG: Traffic Topology Scene Graph for Topology Reasoning in Autonomous Driving
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.
Defines the Traffic Topology Scene Graph (T2SG) and proposes TopoFormer, a one-stage transformer with geometry-guided lane aggregation and counterfactual intervention layers, reaching 46.3 OLS on OpenLane-V2 subsetA.
Problem
Traffic topology reasoning for HD map generation requires modeling lanes, the road signals that control/guide them, and topology relations among lanes; prior methods model lane-lane and lane-signal relations separately or locally and ignore road-structure priors.
Approach
Unifies lanes guided by road signals and their topology into a traffic scene graph, generated by a transformer whose self-attention is biased by inverse geometric distances between lane centerlines (LAL) and trained with a counterfactual zero-attention branch whose total indirect effect supervises edge prediction (CIL).
Key results
- TopoFormer reaches 46.3 OLS on OpenLane-V2 subsetA, versus 44.1 for TopoMLP and TopoLogic.
- On subsetB, TopoFormer reaches 47.5 OLS, versus 42.3 for TopoLogic.
- On T2SG generation, TopoFormer reaches edge accuracy A@1.0 of 8.8, above SGformer (8.5) and the baseline (8.0).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use TopoFormer rather than SGformer | T2SG generation · OpenLane-V2 | A@13.0: 34.6 -> 35.6 | Helped |
| Use TopoFormer rather than TopoLogic | traffic topology reasoning · OpenLane-V2 subsetA | OLS: 44.1 -> 46.3(+2.2) | Helped |
| Use TopoFormer rather than TopoLogic | traffic topology reasoning · OpenLane-V2 subsetB | OLS: 42.3 -> 47.5(+5.2) | Helped |
| Add geometric spatial proximity matrix to lane aggregation (ablation) | TopoFormer · OpenLane-V2 subsetA | TOPll: 21.6 -> 24.1 | Helped |
| Use Hadamard product instead of adding SPM (ablation) | TopoFormer lane aggregation · OpenLane-V2 subsetA | TOPll: 24.1 -> 22.0 | Hurt |
| Use multiplication instead of adding SPM (ablation) | TopoFormer lane aggregation · OpenLane-V2 subsetA | TOPll: 24.1 -> 22.9 | Hurt |
| Add counterfactual intervention layer (ablation) | TopoFormer · OpenLane-V2 subsetA | OLS: 44.9 -> 46.3 | Helped |
| Use zero attention rather than mean attention weights for counterfactual intervention (ablation) | TopoFormer CIL · OpenLane-V2 subsetA | OLS: 45.2 -> 46.3 | Helped |
| Use zero attention rather than random matrices for counterfactual intervention (ablation) | TopoFormer CIL · OpenLane-V2 subsetA | OLS: 45.6 -> 46.3 | Helped |
Surprising findings
- Zero-attention counterfactual intervention (CIL-Zero) performed marginally better than mean or random counterfactual matrices.
- GCN-based scene graph methods (3DSSG, EdgeGCN) drastically lowered edge accuracy relative to the baseline (A@1.0 0.4 vs 8.0).
Limitations
- Gains on T2SG edge accuracy at tighter thresholds are modest, and absolute A@1 values stay low.
- Relies on a geometry-based attention prior that depends on detected centerline accuracy, which the authors note can introduce errors.
Experimental setup
Datasets: OpenLane-V2
Benchmarks: OpenLane-V2 subsetA OpenLane-V2 subsetB
Metrics: AP mAP A@1 DETl DETt TOPll TOPlt OLS
Baselines: 3DSSG EdgeGCN EGTR SGformer STSU VectorMapNet MapTR TopoNet TopoMLP TopoLogic
Builds on: BEVFormer Deformable DETR TopoNet OpenLane-V2 EGTR
Closest papers by meaning
- TopoMLP: A Simple yet Strong Pipeline for Driving Topology Reasoning (ICLR 2024 · similarity 0.93)
- Topo2Seq: Enhanced Topology Reasoning via Topology Sequence Learning (AAAI 2025 · similarity 0.92)
- TopoLogic: An Interpretable Pipeline for Lane Topology Reasoning on Driving Scenes (NeurIPS 2024 · similarity 0.91)
- TopoPoint: Enhance Topology Reasoning via Endpoint Detection in Autonomous Driving (NeurIPS 2025 · similarity 0.90)
- TopoCurve: Geometry-Aware Topology Reasoning via Bézier Curves in Autonomous Driving (NeurIPS 2026 · similarity 0.89)
- LaneSegNet: Map Learning with Lane Segment Perception for Autonomous Driving (ICLR 2024 · 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.