Research map / Detection and segmentation / CNN features
No More Ambiguity in 360deg Room Layout via Bi-Layout Estimation
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 Bi-Layout, a compact 360° room layout model that predicts both enclosed and extended layouts using two learnable global context embeddings fused with image features via a shared guidance module, plus a disambiguate metric.
Problem
Layout annotations for 360° room layout estimation are ambiguous at openings (enclosed vs. extended labels), causing single-prediction models to fail and making evaluation unreliable.
Approach
Invert the usual DETR-style cross-attention so image features are queries and two learnable global context embeddings (one per layout type) are keys/values, letting one shared extractor and guidance module produce two distinct layout predictions with little parameter overhead.
Key results
- On MatterportLayout full test set, disambiguate metric yields 82.57% 3DIoU vs. 81.70% for DOP-Net.
- On the MatterportLayout ambiguous subset, 3DIoU improves from 54.80% (DOP-Net) to 59.97%.
- Ambiguity detection on ZInD achieves precision 0.82 and recall 0.71.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use the disambiguate metric to select between the two layout predictions | Bi-Layout · MatterportLayout full test set | 3DIoU: 82.05 -> 82.57 | Helped |
| use the disambiguate metric to select between the two layout predictions | Bi-Layout · MatterportLayout subset | 3DIoU: 57.08 -> 59.97 | Helped |
| use the disambiguate metric to select between the two layout predictions | Bi-Layout · ZInD-Simple full test set | 3DIoU: 90.25 -> 90.95 | Helped |
| use the disambiguate metric to select between the two layout predictions | Bi-Layout · ZInD-All full test set | 3DIoU: 82.96 -> 84.22 | Helped |
| use the disambiguate metric to select between the two layout predictions | Bi-Layout · ZInD-Simple subset | 3DIoU: 53.76 -> 60.63 | Helped |
| use the disambiguate metric to select between the two layout predictions | Bi-Layout · ZInD-All subset | 3DIoU: 52.78 -> 58.53 | Helped |
| replace add fusion with guided cross-attention using the image feature as query (ablation) | Bi-Layout on MatterportLayout · MatterportLayout | 3D IoU: 81.26 -> 82.57 | Helped |
| replace concatenation fusion with guided cross-attention using the image feature as query (ablation) | Bi-Layout on MatterportLayout · MatterportLayout | 3D IoU: 81.44 -> 82.57 | Helped |
| replace AdaIn fusion with guided cross-attention using the image feature as query (ablation) | Bi-Layout on MatterportLayout · MatterportLayout | 3D IoU: 80.54 -> 82.57 | Helped |
| replace FiLM fusion with guided cross-attention using the image feature as query (ablation) | Bi-Layout on MatterportLayout · MatterportLayout | 3D IoU: 81.39 -> 82.57 | Helped |
| use the image feature as query and global context embedding as key and value instead of the standard reversed QKV (ablation) | Bi-Layout on MatterportLayout · MatterportLayout | 3D IoU: 80.85 -> 82.57 | Helped |
| use a shared feature extractor and guidance module rather than two separate models (ablation) | Bi-Layout on MatterportLayout · MatterportLayout | 3D IoU: 82.72 -> 82.57 | No clear effect |
Surprising findings
- Existing SoTA layout models often predict the opposite layout type from the ground truth at openings, revealing label ambiguity that hurts training and evaluation.
Limitations
- Requires re-annotating enclosed-type labels for MatterportLayout (about 15% of labels) via a semi-automatic procedure.
- Disambiguate metric uses the ground truth to select the better of two predictions, so it is an oracle-style evaluation rather than a deployable one.
- Ambiguity detection is evaluated only on ZInD.
Experimental setup
Datasets: MatterportLayout ZInD
Benchmarks: MatterportLayout ZInD-Simple ZInD-All
Metrics: 2DIoU 3DIoU disambiguate IoU precision recall
Baselines: LED2-Net LGT-Net DOP-Net HorizonNet HoHoNet
Builds on: LGT-Net HorizonNet DETR LED2-Net
Closest papers by meaning
- HouseLayout3D: A Benchmark and Training-free Baseline for 3D Layout Estimation in the Wild (NeurIPS 2025 · similarity 0.89)
- Seg2Reg: Differentiable 2D Segmentation to 1D Regression Rendering for 360 Room Layout Reconstruction (CVPR 2024 · similarity 0.85)
- Direct Numerical Layout Generation for 3D Indoor Scene Synthesis via Spatial Reasoning (NeurIPS 2025 · similarity 0.85)
- OptiScene: LLM-driven Indoor Scene Layout Generation via Scaled Human-aligned Data Synthesis and Multi-Stage Preference Optimization (NeurIPS 2025 · similarity 0.85)
- SceneCraft: Layout-Guided 3D Scene Generation (NeurIPS 2024 · similarity 0.85)
- DeBaRA: Denoising-Based 3D Room Arrangement Generation (NeurIPS 2024 · 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.