Research map / Point clouds / Point cloud classification
BWFormer: Building Wireframe Reconstruction from Airborne LiDAR Point Cloud with Transformer
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 BWFormer, an end-to-end DETR-style Transformer that detects 2D corners in a height map, lifts them to 3D, classifies edges with edge attention, and augments training with LDM-simulated LiDAR scans on Building3D.
Problem
Reconstructing building wireframes from sparse, incomplete, and noisy airborne LiDAR point clouds, where prior methods miss corners and edges and lack labeled data.
Approach
Project the point cloud to a 2D height map, detect 2D corners via heat map, lift them to 3D with Transformer queries initialized from 2D positions plus random height embeddings, and classify candidate edges using edge attention that samples deformable attention along the whole edge; a conditional latent diffusion model simulates LiDAR sampling for augmentation.
Key results
- On Building3D, BWFormer reaches WED 0.238 and ACO 0.204 versus PBWR's 0.271 and 0.222, with corner F1 88.4 vs 81.0 and edge F1 79.4 vs 77.2.
- Ablation: 2D-3D corner detection raises EF1 from 58.2 to 74.7; edge attention adds 3.6 EF1 points; synthetic data gives final EF1 79.4.
- Diffusion-based synthetic LiDAR scanning achieves FID 53.9 and MMD 0.37 (x10^-2), beating uniform sampling baselines (FID 140.3 to 180.2).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace PBWR with BWFormer | Building wireframe reconstruction · Building3D test | WED: 0.271 -> 0.238 | Helped |
| replace PBWR with BWFormer | Building wireframe reconstruction · Building3D test | ACO: 0.222 -> 0.204 | Helped |
| replace PBWR with BWFormer | Building wireframe reconstruction · Building3D test | CF1: +7.4% | Helped |
| replace PBWR with BWFormer | Building wireframe reconstruction · Building3D test | EF1: +2.2% | Helped |
| replace uniform LiDAR scan simulation with conditional latent diffusion scanning (ablation) | Synthetic LiDAR scanning · Synthetic LiDAR scanning | FID: 140.3 -> 53.9 | Helped |
| add 2D-to-3D corner detection (ablation) | BWFormer · Building3D test | WED: 0.463 -> 0.290 | Helped |
| add edge attention (ablation) | BWFormer with 2D-to-3D corner detection · Building3D test | EF1: 74.7 -> 78.3 | Helped |
| add synthetic height-map data augmentation (ablation) | BWFormer with edge attention · Building3D test | WED: 0.253 -> 0.238 | Helped |
| set corner counts to N=100 and NH=200 instead of N=150 and NH=300 (ablation) | BWFormer · Building3D test | WED: 0.253 -> 0.255 | No clear effect |
| set corner counts to N=150 and NH=150 instead of N=150 and NH=300 (ablation) | BWFormer · Building3D test | WED: 0.253 -> 0.322 | Hurt |
| vary edge-attention sampling points from M=3 to M=5 (ablation) | BWFormer · Building3D test | WED: 0.272 -> 0.253 | Helped |
| increase edge-attention sampling points from M=5 to M=7 (ablation) | BWFormer · Building3D test | WED: 0.253 -> 0.259 | Hurt |
Limitations
- Misses corners and edges on extremely sparse point clouds
- Predicts redundant, closely spaced corners and edges due to dense 2D corner prediction
- Ablations on corner count and sampling points trained without synthetic data
Experimental setup
Datasets: Building3D synthetic LiDAR scanning height maps
Benchmarks: Building3D
Metrics: WED ACO corner precision corner recall corner F1 edge precision edge recall edge F1 FID MMD
Baselines: Point2Roof PointMAE PointM2AE Building3D supervised PC2WF PBWR 2.5D dual contouring City3D
Builds on: Deformable DETR Point2Roof PBWR HEAT Latent Diffusion Model
Closest papers by meaning
- PBWR: Parametric-Building-Wireframe Reconstruction from Aerial LiDAR Point Clouds (CVPR 2024 · similarity 0.93)
- LiDAR-Net: A Real-scanned 3D Point Cloud Dataset for Indoor Scenes (CVPR 2024 · similarity 0.86)
- EdgeDiff: Edge-aware Diffusion Network for Building Reconstruction from Point Clouds (CVPR 2025 · similarity 0.85)
- GridFormer: Point-Grid Transformer for Surface Reconstruction (AAAI 2024 · similarity 0.84)
- EdgeMovingNet: Edge-preserving Point Cloud Reconstruction via Joint Geometry Features (CVPR 2025 · similarity 0.84)
- TULIP: Transformer for Upsampling of LiDAR Point Clouds (CVPR 2024 · 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.