Research map / Graph learning / Graph self-supervised learning
Patched Line Segment Learning for Vector Road Mapping
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 PaLiS, a patch-based line-segment representation for vector road graph extraction from satellite images, trained with differentiable rasterization from road masks, reaching SOTA on City-Scale and SpaceNet with 6 GPU hours.
Problem
Vector road mapping from satellite images suffers from ambiguous keypoint representations that make training costly and require extra connectivity networks.
Approach
Divide the image into non-overlapping patches and predict one line segment per road-containing patch, learned via differentiable soft rasterization against road centerline masks; graph connectivity is derived geometrically from segment endpoints and patch types without an extra connectivity network.
Key results
- On City-Scale, the method reaches TOPO F1 79.08 and APLS 68.12 with ResNet-34, the highest among compared methods including RNGDet++ (78.44 / 67.76).
- On SpaceNet, TOPO F1 83.70 and APLS 69.68, exceeding RNGDet++ (82.51 / 67.73).
- Training takes 6 GPU hours on a single V100, reported as a 32-fold reduction versus RNGDet++ (about 192 GPU hours).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use PaLiS with geometric connectivity instead of keypoint-based RNGDet++ | Vector road mapping · City-Scale | TOPO F1: 78.44 -> 79.08 | Helped |
| Use PaLiS with geometric connectivity instead of keypoint-based RNGDet++ | Vector road mapping · SpaceNet | APLS: 67.73 -> 69.68 | Helped |
| Use mask labels rather than unsorted vector labels for line-segment supervision (ablation) | PaLiS · SpaceNet | APLS: 57.34 -> 69.68 | Helped |
| Use mask labels rather than sorted vector labels for line-segment supervision (ablation) | PaLiS · SpaceNet | APLS: 65.28 -> 69.68 | Helped |
| Use geometric connectivity instead of learned relationship connectivity (ablation) | PaLiS graph construction · SpaceNet | APLS: 68.47 -> 69.68 | Helped |
| Set patch size to 8×8 instead of 4×4 (ablation) | PaLiS · SpaceNet | APLS: 67.25 -> 69.68 | Helped |
| Set patch size to 8×8 instead of 16×16 (ablation) | PaLiS · SpaceNet | APLS: 67.58 -> 69.68 | Helped |
| Add the auxiliary pixel-level segmentation branch (ablation) | PaLiS · SpaceNet | APLS: 68.77 -> 69.68 | Helped |
| Set rasterizer sharpness τ_inv to 0.1 instead of 8 (ablation) | PaLiS rasterizer · SpaceNet | APLS: 69.30 vs 69.68 | No clear effect |
| Set rasterizer sharpness τ_inv to 1 instead of 8 (ablation) | PaLiS rasterizer · SpaceNet | APLS: 69.96 vs 69.68 | No clear effect |
| Set rasterizer sharpness τ_inv to 4 instead of 8 (ablation) | PaLiS rasterizer · SpaceNet | APLS: 69.16 vs 69.68 | No clear effect |
| Remove rasterizer projection factor t (ablation) | PaLiS rasterizer · SpaceNet | APLS: 69.68 -> 67.10 | Hurt |
Surprising findings
- Mask-supervised rasterized line-segment learning outperformed direct L1 regression on sorted vector labels (SpaceNet APLS 69.68 vs 65.28).
Limitations
- Patch size is a sensitive hyperparameter; both smaller (4x4) and larger (16x16) patches degrade performance.
- Intersection and overpass (X/T-type) handling relies on hand-designed geometric rules and thresholds.
Experimental setup
Datasets: City-Scale SpaceNet
Benchmarks: City-Scale SpaceNet
Metrics: TOPO precision TOPO recall TOPO F1 APLS
Baselines: DLinkNet Orientation Seg-DLA RoadTracer Sat2Graph TD-Road PaRK-Detect RNGDet RNGDet++
Builds on: DLinkNet PaRK-Detect BoundaryFormer
Closest papers by meaning
- Improved Road Connectivity by Joint Learning of Orientation and Segmentation (CVPR 2019 · similarity 0.89)
- Road Network Representation Learning with the Third Law of Geography (NeurIPS 2024 · similarity 0.86)
- Unveiling the Hidden: Online Vectorized HD Map Construction with Clip-Level Token Interaction and Propagation (NeurIPS 2024 · similarity 0.86)
- LaneSegNet: Map Learning with Lane Segment Perception for Autonomous Driving (ICLR 2024 · similarity 0.84)
- Towards Satellite Image Road Graph Extraction: A Global-Scale Dataset and A Novel Method (CVPR 2025 · similarity 0.84)
- OpenSatMap: A Fine-grained High-resolution Satellite Dataset for Large-scale Map Construction (NeurIPS 2024 · similarity 0.84)
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.