Estimating Surface Reflectance from Images

The Problem: To determine the reflective properties of an object’s surface, such as glossiness, matteness, and roughness, given a single image under unknown, real-world lighting. Motivation: An ability to recognize surface reflectance under uncontrolled conditions provides a valuable cue for inferring the properties or identity of a material, and also allows realistic surface rendering in computer graphics. Figure 1 shows images of nine spheres, each photographed in two real-world settings. The two images of each sphere are completely different at the pixel level because illumination varies from one location to another. Yet, a human observer easily recognizes that the images in each column represent spheres of similar materials, while the images in different columns represent spheres of different materials. A human could classify a sphere photographed in a third setting into one of these nine categories accordingtoitsapparent material properties. We wish to develop acomputer algorithm with a similar abilitytorecognize surface reflectance. illuminations reflectances Figure 1: A reflectance classification problem. Each of nine spheres was photographed under multiple real-world illumination conditions. We trained a nine-way reflectance classifier using the images corresponding to six illuminations, and then used it to classify individual images under the seventh illumination