Computational Polarization 3D: new solution for monocular 3D imaging approach in natural conditions

Abstract

Analyzing the polarimetric properties of the reflected light is a simple, fast, and low-cost approach to recover shape information, but no reported method can accurately recover surface shape from non-Lambertian object in natural conditions with a single view. Another major issue is the lack of effective method to address azimuth ambiguity during the surface reconstruction. To address these two major obstacles, we propose a monocular 3D polarization imaging approach for non-Lambertian surfaces by investigating the subtle difference in specular and diffuse reflection and then separating them using independent component analysis (ICA). Separated pure diffuse reflection enables accurate polarization extraction, based on which a gradient field is derived to determine surface normals. The ambiguity is removed by solving an optimization problem with a reference gradient field. We experimentally demonstrate the performance of the proposed method with man-made object under the controlled illumination and human face under natural illumination.