Research map / Point clouds / Shape and surface completion
CAGE: Continuity-Aware edGE Network Unlocks Robust Floorplan Reconstruction
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 CAGE, an edge-centric end-to-end transformer that reconstructs vector floorplans from point-cloud density maps using a dual-query (perturbed + latent) denoising decoder.
Problem
Vector floorplan reconstruction from noisy, incomplete point-cloud density maps; corner-based polygon representations are fragile to missing corners and line-grouping methods lose fine geometric detail.
Approach
Model each room as a sequence of directed, geometrically continuous wall edges (endpoints need not be vertices) instead of corners, and train a transformer decoder with perturbed (denoising) and latent queries so that edges are refined iteratively into watertight polygons.
Key results
- Achieves 99.1% Room F1, 91.7% Corner F1, and 89.3% Angle F1 on Structured3D, outperforming FRI-Net by +3.9 Corner F1 and +2.4 Angle F1.
- Ablation: edge representation raises RoomFormer-style Room F1 from 96.4% to 98.2%; adding both Edge and DN yields Room F1 99.2% with ResNet-50.
- Cross-dataset (Structured3D to SceneCAD): 85.6% Room IoU and 61.6% Angle F1 without post-processing, competitive with PolyRoom.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace corner-based representation with edge-based representation (ablation) | RoomFormer with ResNet50 · Structured3D | Room F1: 96.4% to 98.2% | Helped |
| replace corner-based representation with edge-based representation (ablation) | RoomFormer with ResNet50 · Structured3D | Corner F1: 85.5% to 89.0% | Helped |
| replace corner-based representation with edge-based representation (ablation) | RoomFormer with ResNet50 · Structured3D | Angle F1: 79.5% to 83.2% | Helped |
| add DN training without edge representation (ablation) | RoomFormer with ResNet50 · Structured3D | Room F1: 96.4 -> 97.1 | Helped |
| add DN training without edge representation (ablation) | RoomFormer with ResNet50 · Structured3D | Corner F1: 85.5 -> 87.3 | Helped |
| add DN training without edge representation (ablation) | RoomFormer with ResNet50 · Structured3D | Angle F1: 79.5 -> 82.2 | Helped |
| combine edge representation and DN training (ablation) | CAGE with ResNet50 · Structured3D | Room F1: 98.2 -> 99.2 | Helped |
| combine edge representation and DN training (ablation) | CAGE with ResNet50 · Structured3D | Corner F1: 89.0 -> 91.3 | Helped |
| combine edge representation and DN training (ablation) | CAGE with ResNet50 · Structured3D | Angle F1: 83.2 -> 87.2 | Helped |
| replace ResNet50 with SwinT V1 while keeping Edge+DN (ablation) | CAGE · Structured3D | Room F1: 99.2 -> 98.9 | Hurt |
| replace ResNet50 with SwinT V1 while keeping Edge+DN (ablation) | CAGE · Structured3D | Corner F1: 91.3 -> 91.7 | Helped |
| replace ResNet50 with SwinT V1 while keeping Edge+DN (ablation) | CAGE · Structured3D | Angle F1: 87.2 -> 88.4 | Helped |
Surprising findings
- Swin Transformer backbones give only marginal or mixed gains over ResNet-50 relative to the Edge and DN modules.
- PolyDiffuse post-optimization brings only marginal improvement to CAGE, suggesting it captures structure without heavy post-processing.
Limitations
- Struggles with dense occlusions and fine structural elements such as interior rooms or columns.
- Output quality varies across checkpoints within the same training run.
- Cross-dataset angle and corner metrics remain well below in-domain results.
Experimental setup
Datasets: Structured3D SceneCAD
Benchmarks: Structured3D SceneCAD
Metrics: Precision Recall F1 IoU
Baselines: Floor-SP HEAT RoomFormer SLIBO-Net FRI-Net PolyRoom PolyGraph PolyDiffuse
Builds on: RoomFormer DN-DETR Deformable DETR BSP-Net
Closest papers by meaning
- MvFFN: Multi-view Floor-Plan Feed-Forward Network for Unposed Wide-Baseline Panorama Layout Reconstruction (NeurIPS 2026 · similarity 0.86)
- HouseLayout3D: A Benchmark and Training-free Baseline for 3D Layout Estimation in the Wild (NeurIPS 2025 · similarity 0.86)
- EdgeMovingNet: Edge-preserving Point Cloud Reconstruction via Joint Geometry Features (CVPR 2025 · similarity 0.85)
- SILSA: Sliding-Window Slice Latents for Topology-Preserving High-Resolution 3D Generation (NeurIPS 2026 · similarity 0.85)
- EdgeDiff: Edge-aware Diffusion Network for Building Reconstruction from Point Clouds (CVPR 2025 · similarity 0.85)
- BADGR: Bundle Adjustment Diffusion Conditioned by Gradients for Wide-Baseline Floor Plan Reconstruction (CVPR 2025 · similarity 0.85)
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.