Researchers diagnose AI shortcut in 3D sensing, cut error 3.3x with PhiCalNet

Single-shot fringe projection profilometry (FPP) is a 3D sensing technique, but learning-based methods have mostly been studied at close range. At long range (over 1m), the inverse-square intensity falloff and ill-posed single-image problem cause standard UNet architectures to exploit an unintended shortcut: they rely on object-boundary shape priors instead of actually decoding fringe phase. Using mechanistic interpretability (linear probing, Grad-CAM) and conformal uncertainty quantification, the team confirmed that this shape-prior shortcut is the dominant source of error. A physics-informed loss penalty alone failed to fix it—the architecture itself was the problem.

Enter PhiCalNet: instead of regressing depth directly, it outputs wrapped phase and applies a fixed differentiable calibration layer that maps phase to depth—removing the shape-prior shortcut from the hypothesis space entirely. On a photorealistic synthetic benchmark of 15,600 images with 50 objects, PhiCalNet cut object MAE by 3.3x to 4.46 mm. The residual error is concentrated in just 0.103% of pixels at the +/-π wrap discontinuity. Pixel-wise conformal UQ further shows that rejecting the top 5% of uncertain pixels reduces RMSE by 64% (20.6→7.4 mm) vs only 3.5% for the baseline, proving the diagnostic convergence of MI and UQ.