Research map / 3D vision / 3D generation
HouseLayout3D: A Benchmark and Training-free Baseline for 3D Layout Estimation in the Wild
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.
Introduces HouseLayout3D, a real-world multi-floor building layout benchmark on Matterport3D, and a training-free pipeline (DN-Splatter mesh, OneFormer skeleton, layout prototype fitting, HovSG room segmentation) that outperforms RoomFormer and SceneScript.
Problem
3D layout estimation models are trained on synthetic single-room or single-floor data and cannot handle complex multi-floor buildings, forcing per-floor processing that loses global context.
Approach
A training-free pipeline that reconstructs a mesh from unposed RGB images, extracts a layout skeleton from semantic segmentation, fits a polygon layout prototype via geometric/connectivity/simplicity losses, and parses it into a building-level scene graph with stairs and doors.
Key results
- On HouseLayout3D, the method reaches Avg F1 0.38 for structures and 0.44 for doors, versus 0.22 for per-floor RoomFormer and 0.26 for per-floor SceneScript.
- On ScanNet++ the method reaches Delta_5 of 67.8 and Delta_10 of 84.7, above SceneScript (55.1, 68.5) and RoomFormer (36.8, 48.9), but with more vertices (83.1 vs 41.2 and 32.5).
- Ablation: removing prototype fitting drops Avg F1 from 0.381 to 0.214.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use HouseLayout3D instead of the strongest reported baseline | 3D layout estimation · HouseLayout3D | Structures Avg F1: 0.26 -> 0.38 | Helped |
| Use HouseLayout3D instead of the strongest reported baseline | 3D layout estimation · HouseLayout3D | Doors Avg F1: 0.28 -> 0.44 | Helped |
| Use HouseLayout3D instead of the strongest reported baseline | 3D layout estimation · HouseLayout3D | Windows Avg F1: 0.15 -> 0.38 | Helped |
| Use HouseLayout3D instead of the strongest reported baseline | 3D layout estimation · HouseLayout3D | Δ5: 22.5 -> 61.1 | Helped |
| Use HouseLayout3D instead of the strongest reported baseline | 3D layout estimation · ScanNet++ DSLR validation | Δ5: 55.1 -> 67.8 | Helped |
| Use HouseLayout3D instead of the strongest reported baseline | 3D layout estimation · ScanNet++ DSLR validation | Δ10: 68.5 -> 84.7 | Helped |
| Add prototype fitting (ablation) | HouseLayout3D · HouseLayout3D | Avg F1: 0.214 -> 0.381 | Helped |
| Add room segmentation (ablation) | HouseLayout3D · HouseLayout3D | Avg F1: 0.359 -> 0.381 | Helped |
Surprising findings
- The per-room baselines do not consistently beat per-floor versions; SceneScript per-floor is better on structures than per-room.
- The training-free method is slower, taking one to two hours per scene versus one to two minutes for the feed-forward baselines.
Limitations
- Runtime of one to two hours per scene on an RTX 4090 is much slower than feed-forward baselines.
- Outdoor elements seen through large windows can introduce artifacts.
- Produces more vertices (less compact) than SceneScript and RoomFormer.
Experimental setup
Datasets: Matterport3D ScanNet++
Benchmarks: HouseLayout3D ScanNet++
Metrics: F1@0.5 Avg F1 Delta_5 Delta_10 #Vertices
Baselines: RoomFormer SceneScript QSlim
Builds on: DN-Splatter HovSG OneFormer Metric3D SceneScript
Closest papers by meaning
- No More Ambiguity in 360deg Room Layout via Bi-Layout Estimation (CVPR 2024 · similarity 0.89)
- OptiScene: LLM-driven Indoor Scene Layout Generation via Scaled Human-aligned Data Synthesis and Multi-Stage Preference Optimization (NeurIPS 2025 · similarity 0.88)
- Unposed Sparse Views Room Layout Reconstruction in the Age of Pretrain Model (ICLR 2025 · similarity 0.88)
- MvFFN: Multi-view Floor-Plan Feed-Forward Network for Unposed Wide-Baseline Panorama Layout Reconstruction (NeurIPS 2026 · similarity 0.87)
- DeBaRA: Denoising-Based 3D Room Arrangement Generation (NeurIPS 2024 · similarity 0.87)
- Direct Numerical Layout Generation for 3D Indoor Scene Synthesis via Spatial Reasoning (NeurIPS 2025 · similarity 0.86)
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.