Repairing Shape-Prior Shortcuts in Long-Range Single-Shot Fringe Projection Profilometry
Direct depth regression in single-shot fringe projection profilometry (FPP) networks suffers from "shape-prior shortcuts," where models learn to predict depth from object boundaries rather than actual fringe phase information. The proposed PhiCalNet architecture forces the network to output a wrapped-phase representation ($\sin\phi, \cos\phi$) and maps it to depth via a fixed, differentiable calibration layer, effectively removing the shortcut architecturally. PhiCalNet achieves a 3.3x reduction
Analysis
TL;DR
- Direct depth regression in single-shot fringe projection profilometry (FPP) networks suffers from "shape-prior shortcuts," where models learn to predict depth from object boundaries rather than actual fringe phase information.
- The proposed PhiCalNet architecture forces the network to output a wrapped-phase representation ($\sin\phi, \cos\phi$) and maps it to depth via a fixed, differentiable calibration layer, effectively removing the shortcut architecturally.
- PhiCalNet achieves a 3.3x reduction in Mean Absolute Error (MAE) compared to UNet baselines, lowering error from 14.54 mm to 4.46 mm on a synthetic benchmark.
- The study introduces pixel-wise conformal uncertainty quantification for FPP, demonstrating that rejecting high-uncertainty pixels at phase discontinuities significantly improves accuracy.
Why It Matters
This research addresses a fundamental failure mode in deep learning-based 3D sensing, showing that increasing model capacity or data volume does not solve shortcut learning if the hypothesis space allows it. By enforcing physical constraints through architecture design rather than loss penalties, it offers a robust path for deploying reliable single-shot profilometry systems in industrial inspection and robotics.
Technical Details
- Problem Identification: Standard UNet baselines plateau at 14.54 mm MAE because they exploit shape priors (object boundaries) instead of fringe phase, a limitation unaddressed by scaling data or parameters.
- PhiCalNet Architecture: The model outputs wrapped phase components $(\sin\phi, \cos\phi)$ and uses a fixed differentiable calibration layer to convert phase to depth, ensuring the solution space adheres to physical optics principles.
- Auxiliary Inputs: To handle the non-injective nature of single-shot mapping, the fringe order is provided as an auxiliary input; sensitivity analysis confirms the model tolerates realistic decoding errors in this parameter.
- Performance Metrics: PhiCalNet reduces object MAE to 4.46 mm (a 3.3x improvement). A three-frame extension further reduces error to 1.16 mm. Residual errors are confined to 0.103% of pixels at $\pm\pi$ wrap discontinuities.
- Uncertainty Quantification: The first application of pixel-wise conformal uncertainty quantification in FPP allows for error localization; rejecting the top 5% of uncertain pixels cuts RMSE by 64%, compared to only 3.5% for the baseline.
Industry Insight
- Architectural Inductive Biases: For physics-based vision tasks, embedding physical laws directly into the network architecture (via differentiable layers) is more effective than relying solely on data-driven optimization or soft loss penalties.
- Reliability in Single-Shot Systems: Single-shot methods are vulnerable to ambiguity; combining them with rigorous uncertainty estimation enables selective rejection of low-confidence predictions, crucial for safety-critical applications.
- Benchmarking Limitations: Standard metrics may mask shortcut learning; interpretability tools and uncertainty quantification are necessary to verify that models are actually solving the intended physical problem rather than exploiting visual correlations.
Disclaimer: The above content is generated by AI and is for reference only.