Single color image photometric stereo for multi-colored surfaces

摘要

We present a photometric stereo method that requires only a single color image. Conventional color photometric stereo methods for single color images cannot deal with multi-colored surfaces, since a color observation at a surface point is insufficient for determining the reflectances and the surface normal at that point. We exploit the global information of surface color and geometry by introducing a surface-color feature that enables classification of a surface into regions of the same color and simultaneously estimate surface normals. The surface-color feature, being invariant in geometry, qualifies the spatial distribution of the square norm of RGB reflectances and attributes surface points of a reflectance norm to the correct color. We discuss the theoretical validity of our surface classification and present a practical algorithm for multi-colored surface recovery. Although some classification ambiguities remain in principle, we show that they can be resolved under a smoothness constraint on the surface geometry. We evaluated the accuracy of our method through simulations and we demonstrated its effectiveness on real scenes.