First Principles of Line Drawings

Abstract

This thesis presents an unsupervised method for creating line drawings from photographs or 3D models. Current methods often rely on high quality paired datasets to automate the creation of line drawings. We observe that line drawings are encodings of scene information that convey 3D shape and semantic meaning. We bake these observations into a set of first principle objectives and train an image translation network to map 3D objects into line drawings. We also explore generation of new styles of line drawings through a novel style confusion loss which averages and combines elements from different styles in a structured manner. User studies and quantitative experiments validate that our method encodes geometry and semantic information into line drawings and improves overall drawing quality.