Problem

Large indoor or mixed scenes can exceed single-machine memory limits. Naive block reconstruction introduces seams, floor ownership becomes ambiguous, textures vary across blocks, and browser delivery suffers when tile size and LOD dependencies are uncontrolled.

Approach

• Use an absolute-scale tile pyramid so every pipeline stage shares the same spatial key and dependency structure.
• Run dense reconstruction at the working tile level with margin-aware streaming cuts to preserve local context.
• Merge neighboring meshes pairwise along known tile boundaries, then carry boundary attributes into visibility, texturing, and simplification.

My Contribution

I designed the overall pipeline architecture around bounded working sets and explicit tile contracts, enabling streamable, memory-controlled processing across all stages.

On the algorithm side, I implemented several core modules:

• Pairwise mesh merging via cut-plane-guided 2D triangulation and mesh hole filling
• Textured mesh simplification with decoupled geometry and attribute resampling
• Global texture color correction formulated as a sparse linear system
• Two-phase LOD generation with shape-aware simplification

I personally wrote more than 70% of the pipeline code, spanning point-cloud partitioning, mesh reconstruction, visibility, texturing, LOD generation, and 3D Tiles delivery.

Outcome

The system supports streamable reconstruction stages and produces LOD-ready 3D Tiles that balance geometry fidelity, load time, and browser interaction.