In this paper, we present a new neural network scheme for3D shape reconstruction. This scheme, named as Smooth Projected Polyhedron Representation Neural Network (SPPRNN), can successively refine polyhedron's vertices parameter of an initial 3D shape based on depth maps of several shaded images taken from multiple views. We considered the depth maps obtained by deploying Tsai-Shah shape-from shading (SFS) algorithm as partial 3D shapes of the object to be reconstructed. The task is to reconstruct a complete representation of a given object from only a limited number of views and erroneous SFS depth maps. We investigate the effectiveness of this scheme through experiment of reconstructing a mannequin object from its images. The recosntruction results is evaluated for each view by comparing mean square error of the depth maps recovered by the proposed scheme and the depth maps recovered by the SFS method alone, in comparison with true depth obtained by 3D scanner. it isexperimentally shown that through hierarchical reconstruction and annealing reinforcement strategies our system gives more exact results. In addition,it corrects and smoothly fuses the erroneous SFS depth-maps.