Research map / 3D vision / 3D generation
PVALane: Prior-Guided 3D Lane Detection with View-Agnostic Feature Alignment
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 PVALane, an end-to-end monocular 3D lane detector that generates sparse prior anchors from front-view features and aligns FV and BEV features through those anchors.
Problem
Monocular 3D lane detection: FV-based anchor methods need dense 3D anchors to offset perspective distortion, causing confusing representations and false positives, while BEV/IPM methods suffer from flat-ground assumptions and loss of semantic context.
Approach
A Prior Anchor Network projects predefined 3D anchors into the FV feature map to score lane objectness and keeps only high-scoring sparse prior anchors; a prior-guided view-agnostic feature alignment module then samples FV and BEV features at these anchors' projections and fuses them.
Key results
- On OpenLane validation with ResNet-50, PVALane reaches 62.7 F1, outperforming BEV-LaneDet by 4.3 and LATR by 0.8 points.
- With ResNet-18 PVALane runs at 108 FPS while reaching 61.2 F1 on OpenLane.
- On ONCE-3DLanes, PVALane reaches 76.35 F1, improving over Anchor3DLane (74.87) by about 1.5 points.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use PVALane with ResNet-50 instead of LATR | OpenLane 3D lane detection · OpenLane validation | F1: 61.9 -> 62.7 | Helped |
| use PVALane with Swin-B instead of PVALane with ResNet-50 | PVALane · OpenLane validation | F1: 62.7 -> 63.4 | Helped |
| use PVALane instead of Anchor3DLane | ONCE-3DLanes 3D lane detection · ONCE-3DLanes validation | F1: 74.87 -> 76.35 | Helped |
| add PAN to the baseline (ablation) | PVALane baseline · OpenLane validation | F1: 55.3 -> 59.3 | Helped |
| add PVFA after PAN (ablation) | PVALane baseline + PAN · OpenLane validation | F1: 59.3 -> 60.1 | Helped |
| add view-specific FV and BEV feature encoding (ablation) | PVALane with PAN and PVFA · OpenLane validation | F1: 60.1 -> 60.5 | Helped |
| set the prior-anchor threshold to 0.2 instead of using no threshold (ablation) | PVALane · OpenLane validation | F1: 60.0 -> 60.5 | Helped |
| raise the prior-anchor threshold from 0.2 to 0.4 (ablation) | PVALane · OpenLane validation | F1: 60.5 -> 60.2 | Hurt |
| raise the prior-anchor threshold from 0.2 to 0.6 (ablation) | PVALane · OpenLane validation | F1: 60.5 -> 58.5 | Hurt |
| add FV features to PVFA (ablation) | PVALane PVFA · OpenLane validation | F1: 59.1 -> 60.5 | Helped |
| add BEV features to PVFA (ablation) | PVALane PVFA · OpenLane validation | F1: 59.8 -> 60.5 | Helped |
Experimental setup
Datasets: OpenLane ONCE-3DLanes
Benchmarks: OpenLane ONCE-3DLanes
Metrics: F1 category accuracy X error (near/far) Z error (near/far) FPS precision recall CD error
Baselines: 3D-LaneNet GenLaneNet PersFormer Anchor3DLane BEV-LaneDet LATR
Builds on: Anchor3DLane PersFormer LaneNet Deformable DETR
Closest papers by meaning
- Rethinking Lanes and Points in Complex Scenarios for Monocular 3D Lane Detection (CVPR 2025 · similarity 0.91)
- DV-3DLane: End-to-end Multi-modal 3D Lane Detection with Dual-view Representation (ICLR 2024 · similarity 0.90)
- LaneCPP: Continuous 3D Lane Detection using Physical Priors (CVPR 2024 · similarity 0.89)
- GLane3D: Detecting Lanes with Graph of 3D Keypoints (CVPR 2025 · similarity 0.89)
- GEONet: Global Enhancement and Optimization Network for Lane Detection (AAAI 2025 · similarity 0.87)
- Sketch and Refine: Towards Fast and Accurate Lane Detection (AAAI 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.