“…When the camera and source positions are known, for example, the surface normal and the surface depth at a specular point are both determined up to a one-parameter family. Constraints of this type have been used computationally for different tasks, including recognition (e.g., [4]) and surface reconstruction (e.g., [5], [6], [7]). In a similar vein, complex illumination environments that are known and controlled have been used to obtain higher order surface information (e.g., curvature) [8], [9], [10], [11], [12], [13], [14], [15] and to extract shape from multiple specular bounces [16].…”