PhiCalNet Fixes Depth Regression Shortcuts in 3D Imaging.

This research addresses a critical limitation in single-shot fringe projection profilometry (FPP) networks: the tendency to exploit "shape-prior shortcuts." Instead of accurately deriving depth from fringe phase, these networks often regress depth directly from object boundaries, leading to significant errors, especially in long-range applications. The proposed solution is PhiCalNet, a novel architecture that outputs a wrapped-phase representation (sinφ, cosφ) and then maps this to depth through a fixed, differentiable calibration layer. This architectural design explicitly removes the shape-prior shortcut, forcing the network to learn the correct physical relationship between fringe patterns and depth. Evaluations on a synthetic benchmark showed that PhiCalNet reduced the object mean absolute error by 3.3 times, achieving 4.46 mm compared to 14.54 mm for the best UNet baseline. The residual errors were primarily confined to phase wrap discontinuities, and a three-frame extension further improved accuracy to 1.16 mm, demonstrating a substantial leap in 3D measurement precision.