Disambiguation of surface normals using single-view Mueller shape-from-polarization

Abstract

Using the information contained in polarimetric measurements to estimate the surface normals of an object or scene is a well-established computer vision task referred to as shape-from-polarization (SfP). Challenges in current physics-based SfP are 1) the assumption of an ideal specular or ideal diffuse polarized bidirectional reflectance distribution function (pBRDF), and 2) the fundamental 180 • ambiguity in surface normal azimuth angle which is present even when an appropriate pBRDF is chosen. In this work, we demonstrate the use of single-view Mueller matrix imaging to address these two challenges. In prior work, Mueller characterization of a material is performed to create a realistic mixed specular and diffuse pBRDF. A depth map is estimated based on the polarimetric measurements and a priori knowledge of the source and camera positions. An unambiguous surface normal is independently estimated from this depth map. Although this surface normal from depth is not expected to be precise, it is used to disambiguate the surface normal azimuth angles. The mean angular error of our single-view Mueller SfP disambiguation method is 22.9 • for a 20cm distance from the source to the center of the object and 34.2 • for 90cm distance for a simulated SNR of 100. This is the first deterministic method known to the authors to disambiguate the surface normals from single-view polarimetric measurements.