Role: Computer Vision Consultant

A 3D spatial awareness pipeline for automated vehicle damage inspection. The system fuses depth and color streams to lift per-pixel defect predictions into the vehicle's body frame, so dents can be detected, localized, and measured in real-world units rather than just flagged in 2D image space.

Why 3D, not 2D

Traditional 2D dent detectors struggle with the same physical defect looking radically different under varying lighting, paint finish, and camera angle. Reflections, decals, and shadows all read as anomalies to a flat image model. By reasoning in 3D, the pipeline recovers the actual surface geometry and isolates true depth deviations from the panel's expected curvature — which dramatically cuts false positives.

Capture and calibration

A handheld capture rig combines a color camera with a depth sensor. Per-frame intrinsics and an extrinsic between the two streams are calibrated so every detected pixel can be back-projected into a metric point cloud. Multiple captures of the same panel are registered together to build a denser surface model than any single frame provides.

Detection in 3D

A 2D segmentation model proposes candidate dent masks per frame; the masks are lifted into the point cloud and compared against a fitted curvature baseline for that panel. Only deviations that survive the 3D geometric check are kept, and each surviving dent is measured for depth, area, and position on the body — outputs an inspector can act on rather than a heatmap they have to interpret.

Technical Highlights:

  • Fused RGB and depth into a calibrated 3D representation of the vehicle surface for sub-millimeter dent localization.
  • Trained a segmentation model on dent imagery and projected its masks into the 3D point cloud to measure dent depth, area, and position on the panel.
  • Built a panel-level curvature baseline so the detector flags only true geometric deviations, suppressing reflection and texture false positives.
  • Tuned the capture and inference loop to run on inspection-station hardware in near real time.