Where Networks Draw Lines: Computer Line Drawing with Deep Learning

Abstract

In this project, we explore a potential new approach for producing computer-generated line drawings from three-dimensional models through the use of deep convolutional neural networks. We base our approach on predicting where people are likely to draw lines to represent a given object, using the dataset collected by Cole et. al [2008]. We evaluate the performance of two network structures, trained on this dataset to take as input, an array of properties representing a three-dimensional object and a viewpoint, and to predict the average of where people would draw lines to represent that object. In addition, we examine two different methods for extracting a clean line drawing from the resulting image.