Research map / Detection and segmentation / CNN features
GSENet:Global Semantic Enhancement 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 GSENet, a CLRNet-based lane detector that adds a Global Feature Extraction Module, a Top Layer Auxiliary Module, an Angle Loss and a Generalized Line IoU loss to improve lane detection in hard scenes.
Problem
Lane detection in challenging scenarios (abnormal weather, occlusion, curves, low visibility) where predicting invisible lanes requires global scene semantics that deep CNN stacks capture inefficiently.
Approach
Combine an MLP-mixer and a dilated multi-head block-wise self-attention branch (with SimAM) into a GEM for global features, distill top-layer global features to classification and regression heads via a TLAM auxiliary head, and add an Angle Loss and a GLIoU loss that penalize angle and point-rectangle mismatch between predicted and GT lanes.
Key results
- On CULane with ResNet18, GSENet reaches F1@50 of 80.42, 0.84 above CLRNet (ResNet18) at 79.58.
- On TuSimple, GSENet (ResNet18) achieves F1 97.98, and reports the lowest FP of 1.79%.
- Ablation on CULane shows Angle Loss, GLIoU, TLAM and GEM each add to F1@50 over the CLRNet baseline, ending at 80.42.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add Angle Loss (ablation) | CLRNet with ResNet18 · CULane | F1@50: 79.58 -> 79.81 | Helped |
| add GLIoU Loss (ablation) | CLRNet with ResNet18 and Angle Loss · CULane | F1@50: 79.81 -> 80.03 | Helped |
| add TLAM (ablation) | CLRNet with ResNet18, Angle Loss, and GLIoU Loss · CULane | F1@50: 80.03 -> 80.31 | Helped |
| add GEM (ablation) | GSENet with ResNet18, Angle Loss, GLIoU Loss, and TLAM · CULane | F1@50: 80.31 -> 80.42 | Helped |
| use 0 TLAM residual blocks (ablation) | TLAM in the ResNet18 model · CULane | mF1: 55.23 -> 55.20 | Hurt |
| use 1 TLAM residual block (ablation) | TLAM in the ResNet18 model · CULane | mF1: 55.23 -> 55.32 | Helped |
| use 2 TLAM residual blocks (ablation) | TLAM in the ResNet18 model · CULane | mF1: 55.23 -> 55.28 | Helped |
| use 3 TLAM residual blocks (ablation) | TLAM in the ResNet18 model · CULane | mF1: 55.23 -> 55.19 | Hurt |
| use 4 TLAM residual blocks (ablation) | TLAM in the ResNet18 model · CULane | mF1: 55.23 -> 55.14 | Hurt |
| set Angle Loss weight to 10 (ablation) | GSENet with ResNet18 · CULane | mF1: 55.23 -> 55.35 | Helped |
| set Angle Loss weight to 15 (ablation) | GSENet with ResNet18 · CULane | mF1: 55.23 -> 55.61 | Helped |
| set Angle Loss weight to 20 (ablation) | GSENet with ResNet18 · CULane | mF1: 55.23 -> 55.12 | Hurt |
Limitations
- Gains on TuSimple are marginal because performance across methods is saturated.
- TLAM number of residual blocks and loss weight show non-monotonic sensitivity, with performance degrading beyond optimal settings.
- The paper reports no inference speed measurements despite claiming faster global feature extraction.
Experimental setup
Datasets: CULane TuSimple
Benchmarks: CULane TuSimple
Metrics: F1@50 F1@75 mF1 accuracy FP
Baselines: CLRNet LaneATT CondLaneNet CANet UFLDv2 RESA SCNN FOLOLane
Builds on: CLRNet UFLD SimAM DropKey MLP-Mixer
Closest papers by meaning
- GEONet: Global Enhancement and Optimization Network for Lane Detection (AAAI 2025 · similarity 0.94)
- A Hybrid Global-Local Perception Network for Lane Detection (AAAI 2024 · similarity 0.92)
- 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)
- Selective Focus: Investigating Semantics Sensitivity in Post-training Quantization for Lane Detection (AAAI 2024 · similarity 0.88)
- LaneSegNet: Map Learning with Lane Segment Perception for Autonomous Driving (ICLR 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.