Semi-automated reconstruction of indoor geometry from 360-degree video for CFD-based airflow analysis in classrooms
TLDR
Presents a semi-automated workflow converting 360-degree videos of rooms into editable simulation-ready geometry assets via NeRF, SAM, ICP, and procedural modeling for CFD airflow analysis.
Reasoning
Strengths include practical integration of modern vision methods with CFD validation and real classroom deployments. Weaknesses are that the workflow is semi-automated and produces simulation geometry rather than fully parametric CAD models, limiting direct relevance to CAD-specific contributions.
Read-first score
Read-first score 46.2, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 30.
Field roles
Rank sensitivity
Stability: volatile; rank range: 40.
Keyword Scores
Deep Analysis
Innovations
- End-to-end semi-automated pipeline converting a single 360-degree video into individually editable, simulation-ready geometry assets
- Use of Neural Radiance Fields for dense point-cloud reconstruction combined with text-prompted SAM 3 instance segmentation lifted to 3D via multi-view consensus and depth-band filtering
- Octree-based object instance separation with connectivity-graph healing of occlusion gaps
- Hybrid object reconstruction using ICP-fitted chair templates and procedurally generated table geometry
- Browser-based editor for quality assurance and rapid reconfiguration of alternative layouts without re-capturing the room
- CFD workflow coupling steady RANS OpenFOAM with transient passive-scalar transport, verified by mesh-sensitivity analysis and validated against IEA Annex 20
Methodology
The workflow reconstructs a dense point cloud from a single 360-degree video using NeRF, then lifts text-prompted SAM 3 2D instance masks to 3D using multi-view consensus and depth-band filtering. Points are separated into object instances with an octree, occlusion gaps are healed with a connectivity graph, chair templates are aligned via ICP, and table geometry is generated procedurally. A browser-based editor supports quality assurance and layout changes, while CFD evaluation uses steady RANS OpenFOAM with transient passive-scalar transport, verified by mesh-sensitivity testing and IEA Annex 20 validation, and applied to two classrooms and a tiered lecture-hall auditorium.
Key Results
The capture-to-geometry pass takes two to five hours per room on a consumer workstation. In a controlled obstruction sequence, modeled half-clearance time varied non-monotonically as furniture was added, and cross-room comparison showed that clearance behavior cannot be reliably extrapolated between rooms.
Limitations
- The workflow is semi-automated rather than fully automated, requiring a browser-based editor for quality assurance and layout construction.
- Geometry acquisition still requires two to five hours per room, limiting very rapid deployment.
- Clearance behavior cannot be reliably extrapolated between rooms, so per-room geometry acquisition remains necessary.