Multi-Criteria Azimuth Correction Method for Monocular Polarization 3D Reconstruction

Polarization 3D imaging has gained attention due to its simple hardware, low cost, and suitability for weakly textured and long-range scenes; however, its inherent 180° azimuth ambiguity restricts practical deployment. This paper proposes a monocular polarization 3D reconstruction method based on multi-criteria azimuth correction, which exploits only the physical and statistical information contained in the polarization images themselves, requiring no external priors or training data. Specifically, Stokes sign consistency is utilized as a physical validity gate, while local continuity and distribution concentration are combined to form a comprehensive scoring function to perform pixel-wise azimuth correction. Subsequently, Frankot–Chellappa frequency-domain integration and Poisson integration are separately employed to recover the 3D surface shape. Experiments on an indoor chair and an outdoor lighthouse demonstrate that the proposed method achieves improved reconstruction quality under both integration frameworks, effectively suppressing depth fluctuations in smooth regions and preserving sharp contours at edges. Quantitative profile fitting and ablation studies confirm its robustness, while its generalizability is validated within the two tested representative scenes.