Using Specularities to Recover Multiple Light Sources in the Presence of Texture

Abstract

Recovering multiple point light sources from a sparse set of photographs in which objects of unknown texture can move is challenging. This is because both diffuse and specular reflections appear to slide across surfaces. What is seldom demonstrated, however, is that it can be taken advantage of to address the light-source recovery problem. In this paper, we therefore show that, if 3D models of the moving objects are available or can be computed from the images, we can solve the problem without any a priori constraints on the number of sources, on their color, or on the surface albedos. Our approach involves finding local maxima in individual images, checking them for consistency across images, retaining the apparently specular ones, and having them vote in a Hough-like scheme for potential light source directions. The precise directions of the sources and their relative power are then obtained by optimizing a standard lighting model. As a byproduct we also obtain an estimate of various material parameters such as the unlighted texture and specular properties. We show that the resulting algorithm can operate in presence of arbitrary textures and an unknown number of light sources of possibly different unknown colors. We also estimate its accuracy using ground-truth data