An evidence combining approach to shape-from-shading

We describe a statistical framework for shape-from-shading. First, we commence by making least squares estimates of local Hessian matrices using samples of surface normals. These estimated Hessians are used to perform parallel transport of the surface normals across the surface. By transporting neighbouring surface normals in this way we are able collect samples of votes for the surface normal direction at each location. We select between these alternatives on the basis of compliance with the image irradiance equation.