Automatic 3D object reconstruction from a single image

Abstract

In this paper, we describe a fully automatic, unsupervised system for recognizing and reconstructing various classes of 3D objects in single-object, light-background 2D images, using a three-layer back-propagation neural network for object classification. Our system takes a single 2D image as input and provides a correctly textured 3D VRML, X3D, or WebGL file for two classes of objects: boxes and spheres. Our approach applies edge detection to the 2D input image, finds the center of gravity of the foreground object, and gathers a set of perimeter distances around the center of gravity. These values are passed to a trained back-propagation neural network with 36 input nodes, 100 intermediate nodes, and four output nodes corresponding to four classes of objects: rectangular prisms (boxes), spheres, cylinders, and organic/other. The approach then deconstructs the object in 2D, reconstructs it in 3D and outputs a textured model in VRML, X3D or WebGL.