Polarization-based capture recovers material parameters of layered dielectrics, but most pipelines report only point estimates. We introduce a Bayesian inverse-rendering framework that yields posteriors and pixelwise credible intervals for surface roughness r, surface–subsurface balance β, absorption a, and refractive index n. A bounded-support likelihood for DOP, heavytailed residuals for intensity, hierarchical priors, and optional spectral coupling ensure stability across geometries and wavelengths. We derive identifiability diagnostics (Fisher information, profile likelihoods) and validate uncertainty via simulation-based calibration and empirical coverage. On synthetic data and real captures (leaves, fruit, plastics/coatings), our method preserves accuracy, meaningfully uses grazing angles, and exposes regimes where parameters are weakly identifiable—insights invisible to point estimates.
Polarization-based capture recovers material parameters of layered dielectrics, but most pipelines report only point estimates. We introduce a Bayesian inverse-rendering framework that yields posteriors and pixelwise credible intervals for surface roughness r, surface–subsurface balance β, absorption a, and refractive index n. A bounded-support likelihood for DOP, heavytailed residuals for intensity, hierarchical priors, and optional spectral coupling ensure stability across geometries and wavelengths. We derive identifiability diagnostics (Fisher information, profile likelihoods) and validate uncertainty via simulation-based calibration and empirical coverage. On synthetic data and real captures (leaves, fruit, plastics/coatings), our method preserves accuracy, meaningfully uses grazing angles, and exposes regimes where parameters are weakly identifiable—insights invisible to point estimates.
