Computational vision as inverse optics: reconstructing spectral reflectances and illuminant

Abstract

The standard definition of computational vision is that it is inverse optics. The direct problem - the problem of classical optics - or computer graphics - is to determine the images of three-dimensional objects. Computational vision is confronted with the inverse and ill-posed problems of recovering surface properties from the partial information contained in images. As a consequence, vision must rely on natural constraints, that is, general assumptions about the physical world to derive an ambiguous output.