Two-source surface reconstruction using polarisation

Abstract

Polarisation vision aims to capture and interpret the polarisation state of incoming light as part of a computer vision system. In this paper, polarisation information is fused with two-source photometric stereo (PS) in order to determine the three-dimensional geometry of surfaces in the presence of both specular and diffuse reflection. In addition to the primary benefit of applying to both reflection types, the method lessens the effects of the Lambertian assumption, which is prevalent in many PS methods other than those requiring many light sources. Further, the method overcomes inherent ambiguities that are present in basic polarisation vision systems. The proposed method commences by using PS to deduce a constrained mapping of the surface normal at each point onto a 2D plane. Phase information from polarisation is used to deduce a mapping onto a different plane. The paper then shows how the full surface normal can be obtained from the two mappings. The results section of the paper demonstrates strong performance of the novel approach against the baseline methods for a range of real-world objects.