Shape Recovery Using Physical Models of Reflection and Interreflection

Our initial effort is directed towards developing a general reflectance framework for image intensity understanding. There are two approaches to the study of reflection: physical and geometrical optics. By studying existing reflectance models based on both geometrical and physical optics, we propose a unified reflectance model that has three primary components: the diffuse lobe, the specular lobe, and the specular spike. The unified model describes the reflection of monochromatic light from surfaces that vary from smooth to rough. By making surface smoothness assumptions, the unified reflectance model may be reduced to the hybrid model; a linear combination of diffuse and specular components. Using the hybrid model, we develop a method that extracts the shape and reflectance of surfaces, simultaneously. The method can therefore adapt to variations in hybrid reflectance from one surface point to the next. This new technique, called photometric sampling, uses an array of extended sources to ensure the detection of both diffuse and specular reflections. The image intensities recorded at each surface point are used to compute local estimates of surface orientation and reflectance parameters. Experiments are conducted on a variety of real surfaces, including, matte painted, smooth metallic, and plastic surfaces. The last part of this thesis addresses the interreflection problem. Previous shape-from-intensity methods assume that points on the surface are only illuminated by the source. This assumption, however, is valid only for convex surfaces; in the case of concave surfaces, points reflect light between themselves. As a result of these interreflections, shape-from-intensity methods produce erroneous shape estimates. We show that for Lambertian surfaces the erroneous estimates have useful invariance properties. Using a physical interreflection model, we develop an iterative algorithm that recovers the actual shape and reflectance of the surface from the erroneous estimates. Experimental results demonstrate the robustness and accuracy of the algorithm. This is the first vision algorithm that recovers accurate shape information in the presence of interreflections. (Abstract shortened with permission of author.).