Data-Driven Sketch Beautification With Neural Feature Representation

In this paper, we present a data-driven approach for beautifying freehand sketches. Our key premise is that the artist-drawn vector shapes equip visual merits for sketching visually appealing shapes, such as clean local shapes and better global structure relationships such as symmetry. However, it is hard to characterize these merits as general rules across different object categories. To overcome this issue, we leverage a neural network feature representation to capture local and global merits across different object categories, and we design our method around this feature representation. We first match the correspondences between sample points on the input sketches and the collected vector shapes using the extracted feature representations. We then design an optimization that encourages the deformed sketch to resemble the vector shape correspondents in the representation space but keeps its original semantic meaning and style. We demonstrate our method on sketches across different shape categories.