Research map / Detection and segmentation / CNN features
Sketch and Refine: Towards Fast and Accurate 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 SRLane, a lane detector that sketches lane proposals from a per-pixel local direction map and refines them with a Lane Segment Association Module and adaptive multi-level feature sampling, reaching 78.9% F1 on CULane at 278 FPS.
Problem
Lane detection requires precise lane shape localization, but proposal-based methods lack flexibility in lane representation and keypoint-based methods need complicated, error-prone post-processing; anchor-based methods also need huge numbers of anchors.
Approach
Estimate a local direction map with a single lightweight conv layer, turn each point and its direction into a straight-line lane proposal (a "sketch", about 40 proposals), then refine proposals with segment-level cross-attention (LSAM, supervised by matched GT) and adaptive Gaussian-weighted multi-scale feature sampling instead of an FPN.
Key results
- SRLane reaches 78.9% F1 on the CULane test set at 3.6 ms latency (about 278 FPS) with ResNet18 on a Titan Xp.
- On Tusimple, SRLane achieves 96.85% accuracy and 97.66% F1, with the lowest latency (3.6 ms) among compared methods, 2.2x faster than CLRNet.
- Ablation on CULane: adaptive sampling adds 0.6 F1 and segment association adds 0.8 F1 over the 85.7 baseline, and combining both gives 87.2 F1 on the CurveLanes subset used for ablation.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use direction-based proposal initialization instead of anchor-based initialization (ablation) | SRLane · CurveLanes subset | detection performance | Helped |
| add adaptive multi-level feature sampling (ablation) | SRLane · CurveLanes subset | F1: $85.7$ -> $86.3$ | Helped |
| add adaptive multi-level feature sampling (ablation) | SRLane with segment association · CurveLanes subset | F1: $86.5$ -> $87.2$ | Helped |
| add Lane Segment Association Module (ablation) | SRLane without adaptive sampling · CurveLanes subset | F1: $85.7$ -> $86.5$ | Helped |
| add Lane Segment Association Module (ablation) | SRLane with adaptive sampling · CurveLanes subset | F1: $86.3$ -> $87.2$ | Helped |
Surprising findings
- The sketched lanes alone reach about 93% accuracy on Tusimple, before refinement.
Limitations
- The paper reports that the approach is designed for 2D lane detection and is only suggested to transfer to 3D lane detection.
- Direction map resolution is fixed at 4x10, so the proposal count is fixed at 40 regardless of input resolution.
Experimental setup
Datasets: Tusimple CULane CurveLanes
Benchmarks: CULane Tusimple CurveLanes
Metrics: F1 accuracy FP rate FN rate latency (ms) FPS parameters MACs
Baselines: LaneAF FOLOLane RCLane GANet SGNet CLRNet UFLDv2 CondLane LaneATT BézierLane
Builds on: CLRNet LaneATT Feature Pyramid Network focal loss DETR-style RoI head
Closest papers by meaning
- GEONet: Global Enhancement and Optimization Network for Lane Detection (AAAI 2025 · similarity 0.91)
- LaneSegNet: Map Learning with Lane Segment Perception for Autonomous Driving (ICLR 2024 · similarity 0.90)
- A Siamese Transformer with Hierarchical Refinement for Lane Detection (NeurIPS 2024 · similarity 0.90)
- GSENet:Global Semantic Enhancement Network for Lane Detection (AAAI 2024 · similarity 0.89)
- A Hybrid Global-Local Perception Network for Lane Detection (AAAI 2024 · similarity 0.89)
- Rethinking Lanes and Points in Complex Scenarios for Monocular 3D Lane Detection (CVPR 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.