Perceiving Shape from Surface Contours via Artificial Neural Networks

Abstract

This thesis explores the challenge of teaching a machine how to perceive shape from surface contour markings. Such markings are commonly used in clothing, data visualizations, and other man-made constructs, because humans have an apparently natural ability to interpret them. By glancing at a simple collection of curves drawn upon a 3D surface, we can quickly glean general shape and curvature information; and such contours drawn on a 2D surface can give the illusion of curvature where there is none. Machines have no such visual intuition, and therefore are not particularly well-equipped to interpret things designed to leverage this human ability. We approach this problem by synthesizing a new dataset of surface grid- and line- marked 3D surfaces (SurfaceGrid) and training a deep neural net to estimate their shape. Our algorithm successfully reconstructs shape from synthetic 3D surfaces rendered with a variety of grid- and line-contour markings with < 0.5% mean-squared relative error, and extracts general shape and curvature information from 2D pictures of 3D mesh models and real-world wireframe objects.