Research map / Detection and segmentation / CNN features
GEONet: Global Enhancement and Optimization 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 GEONet, a lane detection network with a Global Feature Extraction Module, Top-Tier Supplementary Module, whitening BN/contrastive pretraining, and GRIoU and Angle losses; the paper has been retracted at the authors' request.
Problem
Lane detection under occlusion, ambiguous markings, and varying lighting, where existing CNN-based methods struggle to integrate global semantics with local features while keeping inference fast.
Approach
Adds a dual-branch global feature module (MLP-mixer plus dilated attention with SimAM) and a top-level self-attention supplementary module on a CLRNet-style detector, with whitening BN, whitening contrastive pretraining, and geometric losses (rectangle-based GRIoU and angle loss).
Key results
- On CULane with DLA34, GEONet reaches F1@50 of 82.18 versus 80.47 for CLRNet (DLA34).
- With ResNet18, GEONet reaches F1@50 of 81.65, 2.07 points above CLRNet ResNet18 (79.58).
- On TuSimple, GEONet (ResNet18) reaches F1 of 98.21, reported as 0.32 points above the prior leader.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add GFEM (ablation) | GEONet · CULane | F1@50: 81.64 -> 81.81 | Helped |
| add WCL and WBN (ablation) | GEONet · CULane | F1@50: 80.96 -> 81.81 | Helped |
| add TTSM and GFEM (ablation) | GEONet · CULane | F1@50: 80.96 -> 81.81 | Helped |
| use only Angle Loss instead of the integrated loss components (ablation) | GEONet · CULane | F1@50: 81.81 -> 80.47 | Hurt |
| use 2 residual blocks in TTSM (ablation) | GEONet · CULane | F1@50: 80.44 -> 80.69 | Helped |
| use 3 or 4 residual blocks in TTSM (ablation) | GEONet · CULane | F1@50: 80.69 -> 80.45; 80.69 -> 80.49 | Hurt |
| replace IoU Loss with GRIoU Loss (ablation) | GEONet · CULane | F1@75: 61.61 -> 63.44 | Helped |
| set Angle Loss weight to 15 rather than 0 (ablation) | GEONet with GRIoU Loss · CULane | F1@75: 62.89 -> 63.44 | Helped |
| use GEONet with ResNet34 rather than CLRNet with DLA34 | CULane · CULane | F1@50: 80.47 -> 83.07 | Helped |
| use GEONet rather than CLRNet | TuSimple · TuSimple | F1(%): 97.89 -> 98.21 | Helped |
Surprising findings
- The paper was retracted at the authors' request because a comparison appendix with their prior AAAI 2024 work was omitted in the camera-ready version, raising originality concerns.
Limitations
- Gains on TuSimple are small because performance is near saturation.
- Ablation on the number of residual blocks shows performance below baseline with 0 or 1 blocks, indicating sensitivity to design choices.
Experimental setup
Datasets: CULane TuSimple
Benchmarks: CULane TuSimple
Metrics: F1@50 F1@75 mF1 Accuracy FP FN
Baselines: CLRNet CANet CondLane LaneATT UFLD UFLDv2 RESA FOLOLane
Builds on: CLRNet Feature Pyramid Network SimAM W-MSE contrastive learning
Closest papers by meaning
- GSENet:Global Semantic Enhancement Network for Lane Detection (AAAI 2024 · 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 0.91)
- LaneSegNet: Map Learning with Lane Segment Perception for Autonomous Driving (ICLR 2024 · similarity 0.89)
- GLane3D: Detecting Lanes with Graph of 3D Keypoints (CVPR 2025 · similarity 0.88)
- Rethinking Lanes and Points in Complex Scenarios for Monocular 3D Lane Detection (CVPR 2025 · similarity 0.88)
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.