Consumer LiDAR scanning is easy; converting noisy, incomplete phone scans into engineering-grade building models is not. In this talk, we share what we learned converting 150M+ square feet of iPhone/iPad scans into clean, structured geometry with semantic labels. We’ll break down the pipeline: from mesh enhancements to semantic image-based segmentation to global extraction of schematic primitives. We’ll compare classical ML segmentation with newer foundation-model approaches for semantic labeling and show where each fails on real-world imperfect data. Finally, we’ll cover a production QA strategy that utilizes human review to hit accuracy thresholds at scale. Attendees will leave with concrete techniques for building reliable scan-to-CAD/BIM systems using consumer sensors.

