Detection of specularity using color and multiple views

Abstract

This paper presents a model and an algorithm for the detection of specularities from Lambertian reflections using multiple color images from different viewing directions. The algorithm, called spectral differencing, is based on the Lambertian consistency that color image irradiance from Lambertian reflection at an object surface does not change depending on viewing directions, but color image irradiance from specular reflection or from a mixture of Lambertian and specular reflections does change. The spectral differencing is a pixelwise parallel algorithm, and it detects specularities by color differences between a small number of images without using any feature correspondence or image segmentation. Applicable objects include uniformly or nonuniformly colored dielectrics and metals, under extended and multiply colored scene illumination. Experimental results agree with the model, and the algorithm performs well within the limitations discussed.