Three-dimensional shape reconstruction from a single image based on feature learning

Abstract

Many previous works have proposed methods for reconstructing a three-dimensional (3D) shape from a single image. Some of the methods reconstruct a 3D shape using machine learning. These methods learn the relationship between a 3D shape and a 2D image. However, they cannot learn the desirable features of 2D images for 3D reconstruction, because they use only predefined features. Therefore, this paper presents a method for reconstructing the 3D shape by learning features of a 2D image. This method reconstructs a 3D shape by using Convolutional Neural Network (CNN) for feature learning. The pooling layer and the convolutional layer of the CNN enable us to acquire spatial information of an image and automatically select the valuable feature of the image. From the experimental results using human face images, this method can reconstruct the 3D shape with better accuracy than the previous methods.