Research map / 3D vision / 3D generation
Behind the Veil: Enhanced Indoor 3D Scene Reconstruction with Occluded Surfaces Completion
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 a coarse-fine octree feature method with a dual-decoder (scene-specific Geo-decoder and cross-scene 3D Inpainter) that reconstructs visible and occluded indoor surfaces from depth sequences only.
Problem
Indoor 3D scene reconstruction from depth sequences typically recovers only visible surfaces, leaving occluded surfaces (e.g., furniture backsides, floor under furniture) incomplete.
Approach
Represent visible geometry with fine-level octree features decoded by a scene-specific Geo-decoder optimized online, and occluded geometry with coarse-level octree features decoded by a 3D Inpainter pre-trained offline across scenes with complete meshes and frozen at test time.
Key results
- On 3D-CRS, the method reaches 89.7 F1 in ablation full model (Table 4) versus 83.7 for BNV-Fusion at all frames in Table 3, with completeness 89.3 vs 76.5.
- On iTHOR, completeness is roughly 88-90 across six scenes versus about 67-76 for BNV-Fusion (Table 2).
- Completeness stays near 88 when using only 50 depth frames per scene, while BNV-Fusion drops to 62.3 (Table 3).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use the proposed method instead of BNV-Fusion | 3D scene reconstruction · 3D-CRS | completeness: BNV-Fusion: 81.7, 81.2, 74.9, 68.1, 76.6; Ours: 92.3, 90.0, 89.0, 83.3, 92.0 | Helped |
| use the proposed method instead of BNV-Fusion | 3D scene reconstruction · iTHOR | completeness: BNV-Fusion: 71.6, 73.6, 71.9, 70.6, 69.9, 76.2; Ours: 90.1, 92.6, 88.3, 89.7, 88.8, 89.4 | Helped |
| add a 3D Inpainter (ablation) | Ours · 3D-CRS | completeness: 82.2 -> 90.1 | Helped |
| train the 3D Inpainter on cross-scene data (ablation) | Ours · 3D-CRS | completeness: 82.4 -> 90.1 | Helped |
| use both coarse and fine features instead of only coarse features (ablation) | Ours · 3D-CRS | completeness: 88.5 -> 90.1 | Helped |
| use both coarse and fine features instead of only fine features (ablation) | Ours · 3D-CRS | completeness: 82.2 -> 90.1 | Helped |
| vary the number of depth frames per scene (ablation) | Ours · 3D-CRS | completeness: 87.5, 88.1, 88.4, 89.3 | No clear effect |
Surprising findings
- Accuracy of the method is sometimes lower than BNV-Fusion because errors in predicted occluded regions lower the average accuracy.
Limitations
- Accuracy lower than BNV-Fusion in some scenes due to occluded-region prediction errors.
- Relies on complete 3D meshes for cross-scene inpainter training; ScanNet evaluated only qualitatively.
- Dataset release pending Bosch license approval.
Experimental setup
Datasets: 3D-CRS iTHOR ScanNet
Benchmarks: 3D-CRS iTHOR
Metrics: Accuracy Completeness F1
Baselines: TSDF Fusion Go-Surf BNV-Fusion
Builds on: NGLOD BNV-Fusion MonoSDF
Closest papers by meaning
- IndoorGS: Geometric Cues Guided Gaussian Splatting for Indoor Scene Reconstruction (CVPR 2025 · similarity 0.86)
- PlanarSplatting: Accurate Planar Surface Reconstruction in 3 Minutes (CVPR 2025 · similarity 0.86)
- O^2-Recon: Completing 3D Reconstruction of Occluded Objects in the Scene with a Pre-trained 2D Diffusion Model (AAAI 2024 · similarity 0.85)
- Shading Meets Motion: Self-supervised Indoor 3D Reconstruction Via Simultaneous Shape-from-Shading and Structure-from-Motion (CVPR 2025 · similarity 0.85)
- PLANA3R: Zero-shot Metric Planar 3D Reconstruction via Feed-forward Planar Splatting (NeurIPS 2025 · similarity 0.85)
- Total-Decom: Decomposed 3D Scene Reconstruction with Minimal Interaction (CVPR 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.