anya_Gruber
R&D Engineer · Munich
Mar 2026
One thing worth flagging for 2026: the big scan hardware vendors (Zeiss, Hexagon, FARO) are all embedding their own AI inspection modules now, and honestly for standard GD&T checks they're good enough that you don't need to build a custom pipeline unless your defect types are unusual. Hexagon's SFx Asset Management platform has a cloud-based anomaly detection model that learns from your fleet's historical scan data. Where custom ML still wins is when you have application-specific defect types that don't appear in the vendor's training sets — think specialized coatings, exotic alloys, or additive manufactured lattice structures. For those cases the DIY route with Open3D + PyTorch is still the right call.

Then feed deviation_map.png to your trained YOLO model. The full workflow details are in the Open3D docs at open3d.org/docs.

Then feed deviation_map.png to your trained YOLO model. The full workflow details are in the Open3D docs at open3d.org/docs.

Kevin_O'Brien
Quality / NDT Engineer · Cork
Mar 2026
For actual defect detection (not just deviation colormap) there's a pipeline that's getting more traction in aerospace QC — you run the scan deviation data as a 2.5D depth map and feed it through a lightweight CNN trained on your known-defect library. We fine-tuned a YOLOv8 model on annotated deviation maps from about 800 previously inspected parts and it now flags porosity clusters, surface undercuts, and weld underruns with 89% precision on first pass. The open-source starting point is the Ultralytics YOLOv8 library and you can convert deviation maps to usable images with a short Open3D + matplotlib pipeline:

Yuki_Hashimoto
Metrology Engineer · Nagoya
Feb 2026
The scan-to-CAD pipeline has genuinely transformed in the last 18 months because of two things — better ML segmentation of point clouds, and the convergence of photogrammetry and structured light scanning into hybrid tools. In our shop we use a Creaform HandySCAN 3D for capture, pipe the .XYZ into 3D Systems Geomagic Design X for CAD extraction, and then run the deviation analysis in Control X. The AI bit that actually saves time is in the segmentation step inside Geomagic where the software now uses trained surface classification to automatically separate casting surfaces from machined surfaces from gasket faces — something that used to be 2 hours of manual region-growing selection. The accuracy improvement in the ML segmentation between the 2024 and 2025 releases is substantial, surface classification confidence went from about 78% to 94% in our benchmarks on complex pump housings.