3D Face Reconstruction from Single Two-Tone and Color Images

This chapter addresses the problem of reconstructing the three-dimensional shape of faces from single images. We present an algorithm that uses prior knowledge of faces—a single shape model of a face—to eliminate the ambiguities in the reconstruction [13, 14]. The algorithm achieves veridical reconstruction results on images taken under a wide range of viewing conditions. In addition, it can reconstruct the shape of a face from two-tone (“Mooney”) images of faces [15]. Our algorithm demonstrates the importance of “top-down” information in 3D shape reconstruction. The extent to which internal representations affect perception is fundamental to the understanding of cognitive processes. Perceiving the appearance of a 3D shape in an image can be complicated as shapes are distorted by projection and their appearance is affected by lighting as well as by their color and texture. Yet people can readily perceive shape (perhaps qualitatively) merely from one image. A fundamental question therefore is whether the perception of shape is guided primarily by bottom-up processes, in which only image intensities are used along with generic assumptions regarding the statistics of natural scenes, or, alternatively, if it is dictated by top-down processes, which may be driven by memory and attention and