Research map / Point clouds / Shape and surface completion
EdgeDiff: Edge-aware Diffusion Network for Building Reconstruction from 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 EdgeDiff, a conditional diffusion model that generates building wireframes (parameterized midpoint-offset edges) from noise conditioned on point clouds, with an edge attention module, evaluated on Building3D.
Problem
Reconstructing accurate 3D building wireframes from noisy, incomplete, irregularly sampled real-world point clouds, where prior vertex detection and edge regression methods suffer from error accumulation or limited flexibility.
Approach
Reformulates wireframe reconstruction as a noise-to-wireframe conditional diffusion process over midpoint-offset edge parameters, using DDIM sampling with flexible step and edge counts, plus an edge-aware cross-attention guided by a point-wise edge attention predictor.
Key results
- On Building3D Entry-level, EdgeDiff reaches CF1 0.91 and EF1 0.88, versus PBWR at 0.91 and 0.86.
- On the Tallinn subset, EdgeDiff reaches CR 0.74 and ER 0.70, with CF1 0.84 and EF1 0.79, outperforming PBWR (EF1 0.76).
- Midpoint-offset parameterization improves EF1 from 0.19 (endpoint) to 0.88 on Entry-level, and the edge attention supervision raises CR/ER substantially over unsupervised attention.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use EdgeDiff instead of PBWR | Building3D wireframe reconstruction · Entry-level dataset | EF1: 0.86 -> 0.88 | Helped |
| Use EdgeDiff instead of PBWR | Building3D wireframe reconstruction · Tallinn dataset | EF1: 0.76 -> 0.79 | Helped |
| Use EdgeDiff instead of PBWR | Building3D wireframe reconstruction · Tallinn dataset | CF1: 0.80 -> 0.84 | Helped |
| Use endpoint parameterization instead of midpoint-offset (ablation) | EdgeDiff · Entry-level dataset | EF1: 0.88 -> 0.19 | Hurt |
| Use Gaussian padding instead of repeat padding (ablation) | EdgeDiff · Entry-level dataset | EF1: 0.88 -> 0.84 | Hurt |
| Use uniform padding instead of repeat padding (ablation) | EdgeDiff · Entry-level dataset | EF1: 0.88 -> 0.85 | Hurt |
| Set signal scale to 1.0 instead of 2.0 (ablation) | EdgeDiff · Entry-level dataset | EF1: 0.88 -> 0.82 | Hurt |
| Set signal scale to 3.0 instead of 2.0 (ablation) | EdgeDiff · Entry-level dataset | EF1: 0.88 -> 0.85 | Hurt |
| Remove edge attention (ablation) | EdgeDiff · Entry-level dataset | EF1: 0.88 -> 0.83 | Hurt |
| Add edge attention without dedicated supervision (ablation) | EdgeDiff · Entry-level dataset | EF1: 0.88 -> 0.77 | Hurt |
| Use ground-truth edge attention (ablation) | EdgeDiff · Entry-level dataset | EF1: 0.88 -> 0.90 | Helped |
Surprising findings
- Edge attention without supervision degrades recall (CR 0.60, ER 0.66) relative to no edge attention, showing that the attention needs explicit supervision.
Limitations
- Relies on post-processing (NMS and DBSCAN clustering) to produce complete wireframes.
- Edge attention requires ground-truth-derived supervision; unsupervised attention hurts performance.
Experimental setup
Datasets: Building3D Entry-level Building3D Tallinn
Benchmarks: Building3D Entry-level Building3D Tallinn
Metrics: ACO CP CR CF1 EP ER EF1
Baselines: PBWR Point2Roof PC2WF PointMAE PointNet++ Stratified Transformer PointM2AE
Builds on: DDIM DiffusionDet PBWR
Closest papers by meaning
- PBWR: Parametric-Building-Wireframe Reconstruction from Aerial LiDAR Point Clouds (CVPR 2024 · similarity 0.86)
- DiffusionEdge: Diffusion Probabilistic Model for Crisp Edge Detection (AAAI 2024 · similarity 0.86)
- BWFormer: Building Wireframe Reconstruction from Airborne LiDAR Point Cloud with Transformer (CVPR 2025 · similarity 0.85)
- EdgeMovingNet: Edge-preserving Point Cloud Reconstruction via Joint Geometry Features (CVPR 2025 · similarity 0.85)
- CAGE: Continuity-Aware edGE Network Unlocks Robust Floorplan Reconstruction (NeurIPS 2025 · similarity 0.85)
- BADGR: Bundle Adjustment Diffusion Conditioned by Gradients for Wide-Baseline Floor Plan Reconstruction (CVPR 2025 · 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.