Research map / Detection and segmentation / 3D detection and driving
TopoMLP: A Simple yet Strong Pipeline for Driving Topology Reasoning
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 TopoMLP, a query-based pipeline for driving topology reasoning on OpenLane-V2 that pairs a Bézier 3D lane detector and a YOLOv8-assisted traffic detector with MLP lane-lane and lane-traffic topology heads, and adjusts the TOP metric to remove a false-positive loophole.
Problem
Driving topology reasoning from multi-view images requires detecting lane centerlines and traffic elements and then predicting lane-lane and lane-traffic topology; existing topology metrics can be gamed by ranking false positives with high confidence.
Approach
Argues topology performance is bottlenecked by detection, so invests in stronger detectors (Bézier-curve 3D lane queries and YOLOv8-proposal-initialized traffic queries), then uses simple MLP heads with position embedding for topology, plus a correctness-factor modification of the TOP metric.
Key results
- TopoMLP with ResNet-50 and extra YOLOv8 proposals reaches 41.2 OLS on OpenLane-V2 subset_A, versus 35.6 for TopoNet.
- With Swin-B for 48 epochs, TopoMLP reaches 43.7 OLS on subset_A; on subset_B with Swin-B it reaches 44.6 OLS.
- Replacing predicted topology with ground truth gives only small gains (e.g., Swin-B TOP_ll 10.0 and TOP_lt 30.9 with topology GT, only 0.5 and 2.6 above predictions), motivating better detection.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use extra YOLOv8 proposals (ablation) | TopoMLP with ResNet-50 · OpenLane-V2 subset_A | DET_t: 50.0 -> 53.3 | Helped |
| Increase lane queries from 200 to 300 (ablation) | TopoMLP · OpenLane-V2 subset_A | OLS: 36.9 -> 38.2 | Helped |
| Increase lane queries from 300 to 500 (ablation) | TopoMLP · OpenLane-V2 subset_A | OLS: 38.2 -> 38.0 | Hurt |
| Use 3 instead of 4 Bézier control points (ablation) | TopoMLP · OpenLane-V2 subset_A | OLS: 38.2 -> 37.3 | Hurt |
| Use 5 instead of 4 Bézier control points (ablation) | TopoMLP · OpenLane-V2 subset_A | OLS: 38.2 -> 37.1 | Hurt |
| Add lane coordinate embedding to lane-lane topology features (ablation) | TopoMLP · OpenLane-V2 subset_A | TOP_ll: 6.9 -> 7.2 | Helped |
| Add intersection-point L1 supervision (ablation) | TopoMLP · OpenLane-V2 subset_A | TOP_ll: 6.5 -> 7.2 | Helped |
| Include the view transformation matrix in lane features (ablation) | TopoMLP · OpenLane-V2 subset_A | TOP_lt: 21.4 -> 22.8 | Helped |
| Use traffic features instead of only traffic bounding boxes (ablation) | TopoMLP · OpenLane-V2 subset_A | TOP_lt: 22.0 -> 22.8 | Helped |
| Use a stronger Swin-B backbone instead of ResNet-50 | TopoMLP · OpenLane-V2 subset_A | OLS: 38.2 -> 42.2 | Helped |
| Train for 48 instead of 24 epochs with Swin-B | TopoMLP · OpenLane-V2 subset_A | OLS: 42.2 -> 43.7 | Helped |
| Replace predicted topology with ground-truth topology while keeping detections fixed (ablation) | TopoMLP with Swin-B · OpenLane-V2 subset_A | TOP_ll: 9.5 -> 10.0 | Helped |
Surprising findings
- Replacing predicted topology with ground truth yields only minor topology gains, suggesting detection rather than topology reasoning is the main bottleneck.
- Pushing confidence scores to 1.0/0.0 around 0.5 inflates the original TOP metric, e.g., TopoMLP TOP_ll rises from 7.2 to 19.0, exposing a loophole in the metric.
Limitations
- Uses extra precomputed YOLOv8 proposals as an optional component, which adds dependency on an external detector.
- Topology relies on accurate detection; the authors note that missed detections propagate to missed topology connections.
Experimental setup
Datasets: OpenLane-V2
Benchmarks: OpenLane-V2
Metrics: DET_l DET_t TOP_ll TOP_lt OLS
Baselines: STSU VectorMapNet MapTR TopoNet
Builds on: PETR DETR Deformable DETR TopoNet MOTRv2
Closest papers by meaning
- TopoPoint: Enhance Topology Reasoning via Endpoint Detection in Autonomous Driving (NeurIPS 2025 · similarity 0.94)
- TopoLogic: An Interpretable Pipeline for Lane Topology Reasoning on Driving Scenes (NeurIPS 2024 · similarity 0.93)
- T2SG: Traffic Topology Scene Graph for Topology Reasoning in Autonomous Driving (CVPR 2025 · similarity 0.93)
- Topo2Seq: Enhanced Topology Reasoning via Topology Sequence Learning (AAAI 2025 · similarity 0.92)
- TopoCurve: Geometry-Aware Topology Reasoning via Bézier Curves in Autonomous Driving (NeurIPS 2026 · similarity 0.92)
- 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.