level: research
single-shot fringe projection profilometry networks often learn a shortcut: they predict depth from object edges instead of the fringe phase. this happens because regressing depth directly allows the model to ignore the physical measurement principle. on a synthetic benchmark with 15,600 images and 50 objects, a standard unet baseline plateaus at 14.54 mm mean absolute error. adding more data or capacity does not fix the problem, since the shortcut remains a valid solution in the hypothesis space.
the proposed method, phicalnet, avoids the shortcut by design. instead of outputting depth directly, it predicts a wrapped-phase representation (sine and cosine of the phase) and then maps it to depth through a fixed differentiable calibration layer. this architectural change removes the shape-prior solution without needing extra loss terms. because single-shot phase-to-depth mapping is ambiguous without fringe order, the network takes fringe order as an auxiliary input. a sensitivity analysis shows the approach tolerates realistic decoding errors.
the work demonstrates that architectural constraints can enforce physical consistency in learned 3d reconstruction. by forcing the network to reason through phase, the method improves accuracy and generalization. the approach is validated on a photorealistic synthetic dataset with objects at 1.5 to 2.1 meters standoff distance. the design is relevant for industrial inspection and metrology where single-shot accuracy is critical and data efficiency matters.
why it matters: it shows how to build neural networks that respect physical measurement principles, improving reliability in 3d scanning for ai-driven inspection systems.