Research map / Detection and segmentation / CNN features
A Hybrid Global-Local Perception Network for Lane Detection
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 HGLNet, a lane detector that runs a global extraction head (adaptive deformable sampling plus self-attention) in parallel with a local multi-hierarchy feature aggregator (MFA) on a line-anchor lane prior head, reaching SOTA on CULane, LLAMAS and TuSimple.
Problem
Lane detection fails in challenging scenarios such as absent, occluded, faded or blurred lane markings, where existing methods lack sufficient global semantic context and multi-hierarchy local features.
Approach
Decouple global and local feature modeling into two parallel heads: a global extraction head that applies deformable-style adaptive sampling around lane priors and self-attention for global context, and an MFA module that combines dilated convolution and context-aware attention with channel attention to enhance local/regional hierarchies.
Key results
- On CULane, HGLNet (DLA-34) reaches 81.83% F1 at 104 FPS, versus 80.47% F1 at 94 FPS for CLRNet (DLA-34).
- On LLAMAS test, HGLNet (DLA-34) achieves 96.20% F1, improving on LaneATT (ResNet-122) by 2.66 F1 points.
- On TuSimple, HGLNet (ResNet-101) achieves 96.89% accuracy... with the best FP/FN, reported as state-of-the-art accuracy; ablation shows MFA raises CULane mF1 from 52.84 to 54.58.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add MFA (ablation) | HGLNet · CULane | mF1: 52.84% to 54.58% | Helped |
| add MFA (ablation) | HGLNet · CULane Shadow scenario | F1: 79.30 to 80.64 | Helped |
| add MFA (ablation) | HGLNet · CULane No line scenario | F1: 50.16 to 52.90 | Helped |
| add Hybrid architecture to the model with MFA (ablation) | HGLNet · CULane | mF1: 54.58% to 55.83% | Helped |
| use HGLNet instead of CLRNet | CULane test with DLA-34 · CULane test | F1 | Helped |
| use HGLNet instead of LaneATT | LLAMAS test · LLAMAS test | F1: 2.66% F1 | Helped |
| use HGLNet instead of prior methods | TuSimple test · TuSimple test | Accuracy: 96.89% | Helped |
Surprising findings
- Ablation shows the hybrid global head adds about 1.25 mF1 on top of MFA on CULane (54.58 to 55.83), and the authors note the results on TuSimple are close to saturation.
Limitations
- Gains on TuSimple are marginal because performance is near saturation.
- Global head relies on lane prior sampling points, so it is tied to the line-anchor formulation.
Experimental setup
Datasets: CULane TuSimple LLAMAS
Benchmarks: CULane LLAMAS TuSimple
Metrics: F1 (IoU 0.5) mF1 (IoU 0.5-0.95) accuracy FPS FP FN
Baselines: SCNN RESA UFLD LaneATT LaneAF CLRNet PolyLaneNet CondLaneNet
Builds on: CLRNet Deformable DETR LaneATT FOLOLane CondLaneNet
Closest papers by meaning
- GSENet:Global Semantic Enhancement Network for Lane Detection (AAAI 2024 · similarity 0.92)
- GEONet: Global Enhancement and Optimization Network for Lane Detection (AAAI 2025 · similarity 0.92)
- LaneSegNet: Map Learning with Lane Segment Perception for Autonomous Driving (ICLR 2024 · similarity 0.90)
- Sketch and Refine: Towards Fast and Accurate Lane Detection (AAAI 2024 · similarity 0.89)
- A Siamese Transformer with Hierarchical Refinement for Lane Detection (NeurIPS 2024 · similarity 0.88)
- Lane2Seq: Towards Unified Lane Detection via Sequence Generation (CVPR 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.