PhiCalNet slashes 3D depth error 3.3x by repairing shape-prior shortcuts

Traditional single-shot fringe projection profilometry (FPP) networks that directly regress depth often fall into a 'shape-prior shortcut'—they learn to reconstruct depth from object boundaries rather than from the intended fringe phase information. This yields a hard plateau regardless of data or model capacity. In a new arXiv paper, Adam Haroon and colleagues from Iowa State University diagnose this issue and propose PhiCalNet, which architecturally eliminates the shortcut by outputting a wrapped-phase representation (sin φ, cos φ) and mapping it to depth through a fixed differentiable calibration layer. The fringe order is provided as auxiliary input, as the single-shot mapping is non-injective without it, and a sensitivity analysis shows tolerance to realistic decoding errors. On a photorealistic synthetic benchmark of 15,600 fringe images (50 objects at 1.5–2.1 m standoff), PhiCalNet achieves a 3.3× improvement over the best UNet baseline, dropping object MAE from 14.54 mm to 4.46 mm. Residual errors are confined to just 0.103% of pixels at the ±π phase wrap discontinuity. A three-frame extension further reduces error to 1.16 mm, approaching multi-shot accuracy with single-shot capture.

The study also introduces two key validation checks: interpretability analysis confirms phase is the most decodable internal feature, and pixel-wise conformal uncertainty quantification—claimed as the first for FPP—localizes errors at the same wrap discontinuity. Rejecting the top 5% of pixels by snapshot disagreement cuts RMSE by 64% versus only 3.5% for the baseline. The authors stress that the improvement comes from the architectural choice, not from physics-informed loss penalties, as a PINN baseline with identical physics yielded no gain. This work is Part 2 of a two-part study (Part 1: arXiv:2606.17093) and demonstrates that removing shortcuts via architectural design rather than data augmentation can unlock dramatic performance gains in structured-light 3D sensing.