A Neural Network Approach to Shape from Shading

In this paper, we propose a method of recovering shape from shading that solves directly for the surface height using neural networks. The main motivation of this paper is to provide an answer to the open problem proposed by Zhou and Chellappa [11]. We first formulate the shape from shading problem by combining a triangular element surface model with a linearized reflectance map. Then, we use a linear feed-forward network architecture with six layers to compute the surface height with a singular value decomposition. The weights in the model initialized using eigenvectors and eigen-values of the stiffness matrix of objective functional. Experimental results show that our solution is very effective.