Research map / Detection and segmentation / CNN features
A Siamese Transformer with Hierarchical Refinement 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 LATR, a Siamese Transformer that refines lane key-point queries from high-to-low resolution Swin features with shared parameters, plus a Curve-IoU loss for thin, long lane lines.
Problem
Lane detection in complex road conditions (occlusion by vehicles, poor lighting) where local key-point regression is unreliable, and IoU-style losses fail to capture long-distance curved lane misalignment.
Approach
A high-to-low hierarchical refinement Transformer with shared (Siamese) parameters that iteratively refines lane queries over multi-scale Swin features, combined with a Curve-IoU loss that adds an L1-style penalty to the line-IoU so that distant curved predictions receive informative gradients.
Key results
- On OpenLane, the base variant reaches 69.7 F1 and the tiny variant 68.3 F1, 5.0 points above CondLaneNet (ResNet101 best prior) for tiny.
- On CULane, the base variant achieves 80.85 total F1, versus 80.47 for CLRNet with DLA34 (0.38 points over best ResNet-based CLRNet per paper).
- Ablation on CULane: Curve-IoU reaches 80.01 F1 vs 79.24 for LIoU and 77.16 without IoU loss (tiny Swin backbone).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the proposed Siamese Transformer with hierarchical refinement and Curve-IoU | Lane detection model · OpenLane | F1: 59.1 -> 69.7 | Helped |
| Use the proposed Siamese Transformer with hierarchical refinement and Curve-IoU | Lane detection model · CULane | F1: 80.47 -> 80.85 | Helped |
| Use the proposed lane detector | Lane detection model · TuSimple | F1: 97.68 -> 98.01 | Helped |
| Add high-to-low refinement from p0 to p3 instead of refining at p0 only (ablation) | Our method with Swin Transformer tiny · CULane | F1 score: 76.87 -> 80.01 | Helped |
| Add high-to-low refinement from p0 to p3 instead of refining at p3 only (ablation) | Our method with Swin Transformer tiny · CULane | F1 score: 78.15 -> 80.01 | Helped |
| Add Curve-IoU instead of using no IoU loss (ablation) | Lane detection model · OpenLane | F1: 64.28 -> 68.34 | Helped |
| Add Curve-IoU instead of using no IoU loss (ablation) | Lane detection model · CULane | F1: 77.16 -> 80.01 | Helped |
| Add LATR to the LSTR baseline (ablation) | LSTR · OpenLane | F1: 59.96 -> 67.11 | Helped |
| Add LATR to the LSTR baseline (ablation) | LSTR · CULane | F1: 73.56 -> 78.35 | Helped |
| Add Curve-IoU to the LSTR baseline (ablation) | LSTR · OpenLane | F1: 59.96 -> 61.86 | Helped |
| Add Curve-IoU to the LSTR baseline (ablation) | LSTR · CULane | F1: 73.56 -> 76.05 | Helped |
| Increase lane anchors from 50 to 150 (ablation) | Our method · OpenLane | F1: 65.64 -> 68.34 | Helped |
Surprising findings
- Anchor-based methods remain competitive on CULane Shadow, Noline and Arrow scenes, contrary to the OpenLane trend.
- The paper attributes the larger gains on OpenLane to its larger training set, suggesting Transformer methods scale better with data.
Limitations
- Performance drops on some CULane scenes (e.g., Shadow, Noline, Arrow) where anchor-based methods are competitive.
- Gains are smaller on the smaller CULane dataset, suggesting dependence on large training data.
Experimental setup
Datasets: OpenLane CULane Tusimple
Benchmarks: OpenLane CULane Tusimple
Metrics: F1 accuracy FPR FNR FPS GFLOPs
Baselines: LaneATT PersFormer CondLaneNet CLRNet CondLSTR SCNN UFLD LSTR Laneformer RESA PriorLane ADNet
Builds on: CLRNet LSTR Deformable DETR Swin Transformer LaneATT
Closest papers by meaning
- Sketch and Refine: Towards Fast and Accurate Lane Detection (AAAI 2024 · similarity 0.90)
- GSENet:Global Semantic Enhancement Network for Lane Detection (AAAI 2024 · similarity 0.88)
- A Hybrid Global-Local Perception Network for Lane Detection (AAAI 2024 · similarity 0.88)
- GEONet: Global Enhancement and Optimization Network for Lane Detection (AAAI 2025 · similarity 0.88)
- Lane2Seq: Towards Unified Lane Detection via Sequence Generation (CVPR 2024 · similarity 0.87)
- PVALane: Prior-Guided 3D Lane Detection with View-Agnostic Feature Alignment (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.