High-quality polarization 3D reconstruction of weakly textured objects by fusing multi-view images.

Multi-view stereo (MVS) estimates depth by matching features across calibrated views. Though highly accurate on textured surfaces, it is prone to holes and coarse geometry in the case of weak textures or limited viewpoints. Shape-from-polarization (SfP) recovers dense normals from polarized reflections regardless of texture but causes ambiguities in angle inversion, demanding additional disambiguation. Based on these complementary strengths, this paper proposes a passive 3D reconstruction framework that fuses coarse but globally consistent depth priors from a self-attention-enhanced PatchMatch network with fine-grained normal gradients recovered from calibrated polarization measurements. These cues are incorporated into a joint optimization that improves linear consistency between depth and polarization-derived normals while applying a robust, graph-based spatial smoothness constraint to address azimuthal ambiguities and suppress outliers. The final surface is acquired by integrating the optimized normal gradient field in the Fourier domain. Experimental results on different weakly textured objects indicate that our method obtains finer details and fewer artifacts than advanced multi-view stereo and deep learning-based methods, with significantly fewer viewpoints and less computational overhead.