Research map / Detection and segmentation / CNN features
LaneSegNet: Map Learning with Lane Segment Perception for 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 lane segment perception, a unified map representation combining lane geometry, line types and topology, and LaneSegNet, an end-to-end BEV network with lane attention and identical reference-point initialization, evaluated on OpenLane-V2.
Problem
Online HD map learning treats map element detection (geometry only) and centerline perception (topology only) separately, and naive joint prediction of both types of lane is hard to learn and lacks a unified representation.
Approach
Represent each lane as a lane segment (centerline plus left/right boundaries, line types and a directed topology edge), and decode it with a single instance query per segment using a lane attention module (heads-to-regions sampling around multiple reference points) and identical reference-point initialization in the first decoder layer.
Key results
- On lane segment perception, LaneSegNet reaches 32.6 mAP vs 28.5 for the best baseline (MapTRv2), and 14.7 FPS.
- On centerline perception, LaneSegNet achieves 29.7 OLS vs 20.0 for TopoNet, with DET_l 31.8 vs 24.9.
- Ablation: the identical initialization plus heads-to-regions combination raises mAP from 28.8 to 32.6 over the plain decoder, and decoder latency is 31.4% lower than the multi-branch baseline.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use lane segment learning instead of the multi-branch CL + ME formulation (ablation) | LaneSegNet · OpenLane-V2 centerline perception | OLS: +4.8 on OLS | Helped |
| use lane segment learning instead of single-branch CL + ME training (ablation) | LaneSegNet · OpenLane-V2 centerline perception | OLS: +7.2 on OLS | Helped |
| use lane segment learning instead of single-branch CL + ME training (ablation) | LaneSegNet · OpenLane-V2 map element detection | mAP: +4.4 on mAP | Helped |
| add heads-to-regions to the decoder (ablation) | LaneSegNet · OpenLane-V2 lane segment perception | mAP: 28.8 -> 29.1 | Helped |
| add identical reference-point initialization to the decoder (ablation) | LaneSegNet · OpenLane-V2 lane segment perception | mAP: 28.8 -> 30.6; +3.5 | Helped |
| combine heads-to-regions with identical reference-point initialization (ablation) | LaneSegNet · OpenLane-V2 lane segment perception | mAP: 28.8 -> 32.6 | Helped |
| replace lane attention with hierarchical query attention (ablation) | LaneSegNet · OpenLane-V2 lane segment perception | mAP: +0.4 | Helped |
| replace lane attention with hierarchical query attention (ablation) | LaneSegNet · OpenLane-V2 lane segment perception | TOP_lsls: 6.9 -> 5.2 | Hurt |
| use Multi-Point Attention with distributed reference-point initialization (ablation) | LaneSegNet · OpenLane-V2 lane segment perception | mAP: 4.1 decline | Hurt |
| add instance mask supervision (ablation) | LaneSegNet · OpenLane-V2 lane segment perception | AP_ls: +2.0 AP_ls | Helped |
| add instance mask supervision (ablation) | LaneSegNet · OpenLane-V2 lane segment perception | TOP_lsls: +0.6 TOP_lsls | Helped |
| vary the segmentation-loss weight λ_seg (ablation) | LaneSegNet · OpenLane-V2 lane segment perception | results | No clear effect |
Surprising findings
- Identical initialization of reference points (all heads sharing one reference point in the first layer), which seems counter-intuitive, improves mAP by 1.8 over the plain baseline and 3.5 mAP in combination ablations.
Limitations
- Only one backbone (ResNet-50) was tried due to computational limits.
- Not evaluated on nuScenes or Waymo because lane segment annotations are unavailable for them.
- Lane count errors at intersections caused by camera perspective limits.
Experimental setup
Datasets: OpenLane-V2 Argoverse 2
Benchmarks: OpenLane-V2
Metrics: mAP AP_ls AP_ped TOP_lsls AE_type AE_dist OLS DET_l TOP_ll AP_div
Baselines: MapTR MapTRv2 TopoNet STSU VectorMapNet
Builds on: Deformable DETR BEVFormer MapTR TopoNet
Closest papers by meaning
- Sketch and Refine: Towards Fast and Accurate Lane Detection (AAAI 2024 · similarity 0.90)
- TopoLogic: An Interpretable Pipeline for Lane Topology Reasoning on Driving Scenes (NeurIPS 2024 · similarity 0.90)
- A Hybrid Global-Local Perception Network for Lane Detection (AAAI 2024 · similarity 0.90)
- LaneCPP: Continuous 3D Lane Detection using Physical Priors (CVPR 2024 · similarity 0.90)
- Rethinking Lanes and Points in Complex Scenarios for Monocular 3D Lane Detection (CVPR 2025 · similarity 0.90)
- GEONet: Global Enhancement and Optimization Network for Lane Detection (AAAI 2025 · 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.