Research map / Point clouds / Point cloud classification
TopoPoint: Enhance Topology Reasoning via Endpoint Detection 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.
Proposes TopoPoint, which explicitly detects lane endpoints as separate queries, couples them with lanes via attention bias and graph convolutions, and refines endpoints at inference, reaching 48.8 OLS on OpenLane-V2 subset_A.
Problem
Lane endpoints in query-based topology reasoning for autonomous driving deviate from each other because they are supervised through multiple independent lane queries, corrupting lane-lane and lane-traffic topology.
Approach
Independently initialize point and lane queries with explicit endpoint detection, use geometric distances between points and lanes as attention bias in a merged self-attention and point-lane GCNs for mutual feature aggregation, then refine endpoints at inference by geometric matching of detected points to lane endpoints.
Key results
- Achieves 48.8 OLS on OpenLane-V2 subset_A, versus 46.3 for TopoFormer (pretrained lane detector) and 44.1 for TopoLogic.
- Improves endpoint detection DET_p to 52.6 versus 45.2 for TopoLogic on subset_A.
- Reaches 49.2 OLS and 45.1 DET_p on subset_B, state of the art there as well.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace TopoLogic with TopoPoint | OpenLane-V2 subset_A · OpenLane-V2 subset_A | OLS: 44.1 -> 48.8 | Helped |
| Add explicit point detection and point-lane reasoning | OpenLane-V2 subset_A · OpenLane-V2 subset_A | DET_p: 45.2 -> 52.6 | Helped |
| Keep the front-view at original scale (ablation) | TopoPoint · OpenLane-V2 subset_A | DET_t: 46.8 -> 53.8 | Helped |
| Add Point-Lane Merge Self-Attention (ablation) | TopoPoint · OpenLane-V2 subset_A | DET_p: 44.8 -> 49.8 | Helped |
| Add Point-Lane Graph Convolutional Network (ablation) | TopoPoint · OpenLane-V2 subset_A | DET_p: 49.8 -> 51.8 | Helped |
| Add Point-Lane Geometry Matching at inference (ablation) | TopoPoint · OpenLane-V2 subset_A | DET_p: 51.8 -> 52.6 | Helped |
| Add lane-lane GCN (ablation) | Unified Graph Network · OpenLane-V2 subset_A | DET_l: 28.9 -> 29.8 | Helped |
| Add lane-traffic GCN (ablation) | Unified Graph Network with lane-lane GCN · OpenLane-V2 subset_A | DET_l: 29.8 -> 30.6 | Helped |
| Add first Point-Lane GCN (ablation) | Unified Graph Network · OpenLane-V2 subset_A | DET_l: 30.6 -> 30.9 | Helped |
| Add second Point-Lane GCN (ablation) | Unified Graph Network with first Point-Lane GCN · OpenLane-V2 subset_A | DET_p: 51.9 -> 52.6 | Helped |
| Use original-resolution front-view rather than scale 0.5 (ablation) | TopoPoint · OpenLane-V2 subset_A | DET_t: 48.6 -> 55.3 | Helped |
| Downscale multi-view images by a factor of 0.5 (ablation) | TopoPoint · OpenLane-V2 subset_A | DET_l: 30.8 -> 31.4 | Helped |
Surprising findings
- Downscaling multi-view images by 0.5 gave better lane detection than full resolution, while keeping the front view at full resolution was necessary for traffic element detection.
Limitations
- Relies on accurate calibration and may underperform in adverse conditions.
- May underperform in dense traffic scenes due to fixed query settings.
Experimental setup
Datasets: OpenLane-V2
Benchmarks: OpenLane-V2
Metrics: DET_l DET_t TOP_ll TOP_lt OLS DET_p
Baselines: STSU VectorMapNet MapTR TopoNet TopoMLP TopoLogic TopoFormer
Builds on: TopoLogic Deformable DETR BEVFormer GCN
Closest papers by meaning
- TopoLogic: An Interpretable Pipeline for Lane Topology Reasoning on Driving Scenes (NeurIPS 2024 · similarity 0.95)
- TopoMLP: A Simple yet Strong Pipeline for Driving Topology Reasoning (ICLR 2024 · similarity 0.94)
- TopoCurve: Geometry-Aware Topology Reasoning via Bézier Curves in Autonomous Driving (NeurIPS 2026 · similarity 0.92)
- Topo2Seq: Enhanced Topology Reasoning via Topology Sequence Learning (AAAI 2025 · similarity 0.91)
- T2SG: Traffic Topology Scene Graph for Topology Reasoning in Autonomous Driving (CVPR 2025 · similarity 0.90)
- LaneSegNet: Map Learning with Lane Segment Perception for Autonomous Driving (ICLR 2024 · similarity 0.89)
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.