Research map / 3D vision / 3D generation
LaneCPP: Continuous 3D Lane Detection using Physical Priors
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 LaneCPP, a monocular 3D lane detector using a full 3D B-spline lane representation regularized by physical priors (parallelism, surface smoothness, curvature) and a geometry-aware depth-based BEV transformation with learned road surface hypotheses.
Problem
Monocular 3D lane detection needs flexible yet robust lane representations that model complex road structures while behaving predictably, especially given limited and noisy 3D annotations.
Approach
Uses a continuous full-3D B-spline lane representation with control points constrained to normal directions of initial proposals, enabling analytically computed tangent/normal-based soft physical-prior losses (parallelism, surface smoothness, curvature), plus a spatial transformation that lifts front-view features along rays onto multiple road-surface plane hypotheses guided by a surface height lo
Key results
- On OpenLane, LaneCPP reaches 60.3% F1, outperforming BEV-LaneDet (58.4%) and Anchor3DLane (53.7%).
- Prior regularization on OpenLane300 raises F1 from 65.0% to 66.7% with combined priors, and the full model reaches 66.9% (+4.0 over the 62.9% baseline).
- On Apollo 3D Synthetic balanced scenes, LaneCPP achieves the highest F1 of 97.4%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add the novel continuous lane representation (ablation) | modified 3D-SpLineNet baseline · OpenLane300 | F1: 62.9 -> 65.0 | Helped |
| add combined parallelism, surface smoothness, and curvature priors (ablation) | LaneCPP with novel lane representation · OpenLane300 | F1: 65.0 -> 66.7 | Helped |
| use the geometry-aware spatial transformation instead of IPM (ablation) | LaneCPP with novel lane representation · OpenLane300 | F1: 65.0 -> 66.6 | Helped |
| combine the novel lane representation, prior regularization, and spatial transformation (ablation) | modified 3D-SpLineNet baseline · OpenLane300 | F1: 62.9 -> 66.9 | Helped |
| add the parallelism prior (ablation) | LaneCPP without prior losses · OpenLane300 | F1: 65.0 -> 66.2 | Helped |
| add the surface smoothness prior (ablation) | LaneCPP without prior losses · OpenLane300 | F1: 65.0 -> 65.8 | Helped |
| add the curvature prior (ablation) | LaneCPP without prior losses · OpenLane300 | F1: 65.0 -> 66.7 | Helped |
| increase surface hypotheses from 1 to 5 (ablation) | LaneCPP spatial transformation · OpenLane300 | F1-Score: 65.0 -> 66.6 | Helped |
| increase surface hypotheses from 5 to 15 (ablation) | LaneCPP spatial transformation · OpenLane300 | F1-Score: 66.6 -> 66.1 | Hurt |
| increase surface hypotheses from 5 to 27 (ablation) | LaneCPP spatial transformation · OpenLane300 | F1-Score: 66.6 -> 66.0 | Hurt |
| use LaneCPP instead of BEV-LaneDet | OpenLane test · OpenLane test | F1-Score: 58.4 -> 60.3 | Helped |
| use LaneCPP instead of BEV-LaneDet | Apollo 3D Synthetic Rare Scenes · Apollo 3D Synthetic Rare Scenes | F1: 97.6 -> 96.2 | Hurt |
Limitations
- Indicator function may wrongly apply parallelism loss to weakly converging line pairs
- Plane-based surface hypotheses may limit the spatial transformation
- Apollo 3D Synthetic is small and simple, so priors help less there
Experimental setup
Datasets: OpenLane OpenLane300 Apollo 3D Synthetic
Benchmarks: OpenLane Apollo 3D Synthetic
Metrics: F1-Score X-error near X-error far Z-error near Z-error far
Baselines: 3D-LaneNet Gen-LaneNet PersFormer CurveFormer BEV-LaneDet Anchor3DLane GP 3D-SpLineNet
Builds on: 3D-SpLineNet PersFormer LSS (depth-based lift-splat)
Closest papers by meaning
- Rethinking Lanes and Points in Complex Scenarios for Monocular 3D Lane Detection (CVPR 2025 · similarity 0.91)
- LaneSegNet: Map Learning with Lane Segment Perception for Autonomous Driving (ICLR 2024 · similarity 0.90)
- PVALane: Prior-Guided 3D Lane Detection with View-Agnostic Feature Alignment (AAAI 2024 · similarity 0.89)
- GLane3D: Detecting Lanes with Graph of 3D Keypoints (CVPR 2025 · similarity 0.88)
- Sketch and Refine: Towards Fast and Accurate Lane Detection (AAAI 2024 · similarity 0.88)
- GEONet: Global Enhancement and Optimization Network for Lane Detection (AAAI 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.