Diagnosing and Repairing Shape-Prior Shortcuts in Long-Range Single-Shot Fringe Projection Profilometry

Short summary

Researchers diagnose long-range 3D scanning failures using mechanistic interpretability: models exploit shape priors instead of decoding fringe phase. PhiCalNet, an architectural redesign that outputs wrapped phase with fixed calibration, reduces errors 3.3x by enforcing the correct solution path. Conformal uncertainty quantification confirms the diagnosis.

• •Deep learning models for long-range fringe projection fail by using shape priors instead of decoding fringe phase
• •Mechanistic interpretability and conformal UQ identify the same architectural failure mode
• •PhiCalNet reduces errors 3.3x by forcing phase-based inference through architectural design

Generated with AI, which can make mistakes.