Shape from polarization with ordered fusion of physical priors

Shape from polarization (SfP) recovers dense surface normals from a single polarimetric image by exploiting the partial polarization produced when unpolarized light reflects off a dielectric surface. Existing deep SfP methods fuse polarization, shading, and physics priors in a single stage and treat them as equally reliable, even though their underlying assumptions differ in strength. In complex scenes, the weaker priors can corrupt the polarization anchor and degrade reconstruction. We propose CPF-Net, a cascaded prior-fusion network that orders priors by assumption strength and separates disambiguation from refinement into two successive stages. The first stage resolves the azimuth ambiguity by injecting shading cues as a gated residual into the polarization anchor. The second stage refines the zenith angle with Fresnel physical descriptors through multi-scale pyramid fusion. On three real datasets, CPF-Net attains a mean angular error of 14.44° on the challenging SfP-Wild benchmark, a 12.8% reduction over the strongest compared method. It also attains the lowest mean angular error among the compared methods on our self-collected Cpnd dataset, although this margin lies within run-to-run variability, and a slightly lower MAE than the strongest compared method on DeepSfP. Adding only 29 K parameters to the backbone, CPF-Net attains these results with negligible fusion overhead, and the polarization anchor stays dominant where weaker priors fail.