Problem

Raw indoor reconstruction meshes often contain layered walls, floating fragments, holes near corners, and furniture fused into structural surfaces. These artifacts make wall lines ambiguous and break downstream measurement, floorplan, and visualization workflows.

Approach

• Extract and merge dominant planes from the mesh while keeping optional Manhattan alignment as a controlled fallback.
• Project visibility evidence into an occupancy grid to recover free space, floor and ceiling height cues, and room boundaries.
• Combine geometry rules, mask evidence, and graph optimization to separate walls, floors, ceilings, doors, beams, and non-structural objects.

My Contribution

I independently designed and implemented the entire structural reconstruction pipeline from scratch. The five-stage geometry-only processing chain converts raw, noisy multi-layer meshes into closed structural models through robust geometric cues rather than model-dependent recognition, with layered fallbacks at each stage to handle incomplete scans and heavily cluttered rooms.

On the algorithm side, I built the core modules:

• Iterative plane extraction with multi-scale merging on noisy meshes
• Occupancy-grid projection for room topology and floor/ceiling cues
• Constrained Delaunay triangulation with multi-label graph-cut for room refinement
• Incremental cell-complex construction with max-flow surface extraction
• Non-structural object separation via visibility-based filtering

Outcome

The resulting model is cleaner, single-layer, and structurally closed enough for floorplan generation, area estimation, and user-facing VR display.