Research map / 3D vision / 3D generation
GLane3D: Detecting Lanes with Graph of 3D Keypoints
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.
GLane3D detects 3D lane keypoints with multiple proposals and PointNMS, predicts directed connections between them as an adjacency matrix, and extracts lanes from the resulting graph, improving F1 and cross-dataset generalization on OpenLane and Apollo.
Problem
3D lane detection from front-view cameras needs robust generalization to diverse, unseen lane structures, which top-down instance-based methods struggle with.
Approach
Formulate 3D lane detection as locating keypoints and predicting directed connections between them (a graph), with multiple keypoint proposals per lane location to improve recall, PointNMS to remove redundant proposals, and customized BEV anchor geometry for denser near-field IPM sampling.
Key results
- On OpenLane with ResNet-50 backbone, GLane3D-Base reaches 63.9 F1 at 1.5m vs. 62.7 for PVALane (ResNet-50); GLane3D-Large (Swin-B) reaches 66.0 F1 vs. 63.4 for PVALane (Swin-B).
- Cross-dataset (OpenLane-trained, Apollo Balanced Scenes) F1 at 1.5m is 58.9 vs. 34.3 for LATR and 53.2 for PersFormer.
- GLane3D-Lite runs at 62.2 FPS with 61.5 F1 on OpenLane, faster than the compared methods listed.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add PointNMS (ablation) | GLane3D on OpenLane300 · OpenLane300 | F1-score: 66.6 -> 69.2 | Helped |
| Add multiple keypoint proposals without PointNMS (ablation) | GLane3D on OpenLane300 · OpenLane300 | F1-score: 66.6 -> 42.7 | Hurt |
| Add PointNMS together with multiple keypoint proposals (ablation) | GLane3D on OpenLane300 · OpenLane300 | F1-score: 66.6 -> 71.6 | Helped |
| Add custom BEV geometry to PointNMS and multiple keypoint proposals (ablation) | GLane3D on OpenLane300 · OpenLane300 | F1-score: 71.6 -> 72.0 | Helped |
| Increase keypoints from 256 to 512 (ablation) | GLane3D on OpenLane300 · OpenLane300 | F1-score: 72.0 -> 72.4 | Helped |
| Increase keypoints from 256 to 1024 (ablation) | GLane3D on OpenLane300 · OpenLane300 | FPS: 27.8 -> 21.0 | Hurt |
| Use the full GLane3D method instead of PVALane with the same Swin-B backbone | GLane3D-Large · OpenLane test, 1.5m threshold | F1-score: 63.4 -> 66.0 | Helped |
| Use the full GLane3D method instead of PersFormer | GLane3D · Apollo balanced scenes, cross-dataset (OpenLane-trained), 1.5m threshold | F1-score: 53.2 -> 58.9 | Helped |
Surprising findings
- Multiple keypoint proposals alone decreased F1 (from 66.6 to 42.7 on OpenLane300) and only improved performance when combined with PointNMS.
Limitations
- Performance saturates beyond 256 keypoints while FPS drops, indicating a speed-accuracy trade-off.
- Camera+LiDAR fusion is a simple feature concatenation without advanced fusion techniques.
- Multiple proposals alone introduce connection ambiguity that requires PointNMS to resolve.
Experimental setup
Datasets: OpenLane Apollo
Benchmarks: OpenLane Apollo
Metrics: F1-score X-error near/far Z-error near/far FPS
Baselines: PersFormer BEV-LaneDet Anchor3DLane LATR LaneCPP PVALane DV-3DLane M2-3DLaneNet
Builds on: BEVLaneDet LATR PersFormer Deformable attention SECOND
Closest papers by meaning
- Rethinking Lanes and Points in Complex Scenarios for Monocular 3D Lane Detection (CVPR 2025 · similarity 0.90)
- PVALane: Prior-Guided 3D Lane Detection with View-Agnostic Feature Alignment (AAAI 2024 · similarity 0.89)
- Sketch and Refine: Towards Fast and Accurate Lane Detection (AAAI 2024 · similarity 0.89)
- LaneCPP: Continuous 3D Lane Detection using Physical Priors (CVPR 2024 · similarity 0.88)
- Bezier Everywhere All at Once: Learning Drivable Lanes as Bezier Graphs (CVPR 2024 · similarity 0.88)
- GEONet: Global Enhancement and Optimization Network for Lane Detection (AAAI 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.