High-speed and high-accuracy 3D shape recovery is essential for online industrial inspection and robotic perception. Single-shot fringe projection profilometry (FPP) enables instantaneous acquisition, yet suffers from severe absolute-phase/fringe-order ambiguity induced by the \(2\pi \) periodicity. Shape from polarization (SfP) provides dense surface normals and fine local details, but is vulnerable to noise in low-DoLP regions and exhibits intrinsic \(\pi \) -ambiguity in azimuth together with low-frequency geometric drift. To address these complementary failure modes, we propose a Phase–Normal Collaborative Network (PNC-Net) that deeply fuses FPP and SfP under a unified physics-consistency framework. PNC-Net adopts a dual-branch encoder with multi-scale interactions and introduces a bidirectional geometric consistency loss with a symmetric stop-gradient mechanism, enabling cross-teaching between the two modalities: the global phase trend from FPP acts as a geometric anchor to disambiguate polarization normals and suppress low-frequency distortions, while the polarization-derived normals provide high-frequency differential constraints to correct local phase-jump errors and texture-induced artifacts in single-shot phase unwrapping. To ensure robust generalization, we build a large-scale hybrid dataset combining physics-based rendering and real-world acquisition. Extensive experiments demonstrate that PNC-Net achieves state-of-the-art performance on absolute phase, surface normals, and depth, reducing depth RMSE by 39.3% and normal angular error by 38.4% compared with representative single-shot paradigms, while maintaining an inference latency of about 3 ms for real-time 3D recovery.
