Problem

Single-image panoramic depth is fast and low-cost, but independent depth predictions create scale drift, double walls, floating outliers, noisy planes, and holes when fused into a scene-level mesh.

Approach

• Optimize indoor poses with geometric constraints that use room structure rather than relying only on fragile feature matches.
• Filter low-confidence depth and cross-view conflicts before fusion to reduce glass, sky, mirror, and wall leakage artifacts.
• Build mesh cleanup around planar structure and visibility so the final model is more stable for inspection and measurement.

My Contribution

I independently designed and implemented the full refinement workflow for monocular panoramic depth reconstruction.

On the algorithm side, I built the core modules:

• Pose refinement with geometric constraints leveraging room structure rather than fragile feature matches
• Cross-view depth filtering to suppress glass, mirror, and wall leakage artifacts before fusion
• Noise-resistant Graph-Cut mesh reconstruction with BVH-guided invalid region detection
• View-dependent mesh refinement via spherical reprojection for close-range detail recovery

Outcome

The refined output reduces common monocular-depth artifacts and makes panoramic reconstruction more usable for lightweight indoor modeling workflows.