Research map / Point clouds / Point cloud classification
PBWR: Parametric-Building-Wireframe Reconstruction from Aerial LiDAR Point Clouds
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 PBWR, an end-to-end transformer that regresses parameterized 3D building wireframe edges directly from aerial LiDAR roof point clouds, using Hausdorff-based edge similarity for matching, E-NMS, and an edge loss, without corner prediction.
Problem
3D building wireframe reconstruction from sparse aerial LiDAR point clouds; existing methods rely on intermediate corner prediction and edge classification, which accumulate errors and generate many useless edge proposals.
Approach
Directly regress seven edge parameters (midpoint offset, per-axis component lengths, quadrant classes) from transformer decoder queries initialized from farthest-point-sampled points, match predictions to ground truth with an edge similarity combining Hausdorff distance, direction and length, and suppress duplicates with edge-similarity NMS (E-NMS) plus soft-label confidence loss.
Key results
- On Entry-level Building3D, PBWR reaches edge recall 0.82 and EF1 0.86 versus 0.46 and 0.61 for Building3D-supervised.
- On Tallinn city data, PBWR-Tallinn achieves edge recall 0.65 and EF1 0.76 versus 0.23 and 0.36 for Building3D-supervised.
- Ablation shows hard labels without E-NMS or soft edge-similarity labels collapse edge recall to 0.35-0.41, while full model reaches 0.82.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Directly regress parameterized edges rather than edge endpoints (ablation) | PBWR · Building3D Entry-level test | edge recall (ER): 0.12 -> 0.82 | Helped |
| Use Hausdorff edge matching instead of corner matching (ablation) | PBWR · Building3D Entry-level test | edge recall (ER): 0.30 -> 0.82 | Helped |
| Use Hausdorff edge matching instead of midpoint matching (ablation) | PBWR · Building3D Entry-level test | edge recall (ER): 0.81 -> 0.82 | Helped |
| Add E-NMS (ablation) | PBWR · Building3D Entry-level test | edge precision (EP): 0.78 -> 0.91 | Helped |
| Add E-NMS (ablation) | PBWR · Building3D Entry-level test | edge recall (ER): 0.41 -> 0.82 | Helped |
| Add edge similarity loss (Lsim) (ablation) | PBWR · Building3D Entry-level test | edge recall (ER): 0.80 -> 0.82 | Helped |
| Add edge similarity loss (Lsim) (ablation) | PBWR · Building3D Entry-level test | edge precision (EP): 0.89 -> 0.91 | Helped |
| Use edge similarity as soft confidence labels instead of hard labels (ablation) | PBWR confidence-score loss · Building3D Entry-level test | edge recall (ER): 0.35 -> 0.80 | Helped |
| Use edge similarity as soft confidence labels instead of hard labels (ablation) | PBWR confidence-score loss · Building3D Entry-level test | edge precision (EP): 0.56 -> 0.89 | Helped |
| Replace Building3D-supervised baseline with PBWR | Building3D · Entry-level test | edge recall (ER): 0.46 -> 0.82 | Helped |
| Replace Building3D-supervised baseline with PBWR-Tallinn | Building3D · Tallinn test | edge recall (ER): 0.23 -> 0.65 | Helped |
Surprising findings
- PBWR-Corner (directly regressing endpoints) performs poorly (edge recall 0.12), showing endpoint regression is much worse than midpoint-plus-parameter regression.
Limitations
- Generated edges are not a continuously connected set, requiring DBSCAN post-processing to merge corners with a heuristic 0.05 m threshold.
- Heuristic-free design still relies on bipartite matching and tuned balancing coefficients for edge similarity.
- Evaluated only on Building3D roof point clouds (Entry-level and Tallinn).
Experimental setup
Datasets: Building3D Entry-level Building3D Tallinn city
Benchmarks: Building3D Entry-level Building3D Tallinn city
Metrics: ACO corner precision corner recall corner F1 edge precision edge recall edge F1
Baselines: PointNet++ RandLA-Net DGCNN PAConv Stratified Transformer Point2Roof Building3D-supervised PC2WF PointMAE PointM2AE
Builds on: Point2Roof PC2WF Building3D DETR-style transformer decoder
Closest papers by meaning
- BWFormer: Building Wireframe Reconstruction from Airborne LiDAR Point Cloud with Transformer (CVPR 2025 · similarity 0.93)
- EdgeDiff: Edge-aware Diffusion Network for Building Reconstruction from Point Clouds (CVPR 2025 · similarity 0.86)
- LiDAR-Net: A Real-scanned 3D Point Cloud Dataset for Indoor Scenes (CVPR 2024 · similarity 0.85)
- EdgeMovingNet: Edge-preserving Point Cloud Reconstruction via Joint Geometry Features (CVPR 2025 · similarity 0.84)
- Learnable Earth Parser: Discovering 3D Prototypes in Aerial Scans (CVPR 2024 · similarity 0.83)
- STAR-Edge: Structure-aware Local Spherical Curve Representation for Thin-walled Edge Extraction from Unstructured Point Clouds (CVPR 2025 · similarity 0.83)
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.