Three dimensional shape estimation from two dimensional images

In two dimensional images of three dimensional reflecting surfaces, both the shape of the surface and the positions of sources of illumination determine relative values of image intensity. Segmentation of objects and estimation of surface shapes from 2D images are important research areas in scene analysis and computer vision. Many analyses of shape determination assume Lambertian surfaces which have no shadows, diffuse illumination, or specular reflections. While these results provide a good basis for shape from shading analysis, the nonlinear effects specifically omitted from consideration will, in most images of real objects, cause more error in shape estimation than additive white Gaussian noise processes. This paper considers segmentation of objects and estimation of surface shape in the presence of these nonlinear effects and demonstrates results both with well-behaved synthetically generated images and with real images acquired under somewhat controlled conditions.