Reformed Residual Network With Sparse Feedbacks for 3D Reconstruction From a Single Image

Deep neural networks are difficult to train due to the large number of unknown parameters. To increase the trainable performance, we present a novel network model with moderate depth for three-dimensional reconstruction. The proposed network model, called SFResNet, only has eight layers, and sparse feedbacks were added in the middle and last layers, which is mainly used to add the constraints and improve the stability of the network model. In addition, a joint strategy is proposed to reduce the artificial Mosaic trace at the seam of the patches; hence, SFResNet can also evaluate an input image of any size. Visually pleasing output results can be produced with a reconstructed shape and normal surface. The experimental results show the effectiveness of the proposed method.