Keypoint-driven line drawing vectorization via PolyVector flow

Line drawing vectorization is a daily task in graphic design, computer animation, and engineering, necessary to convert raster images set of curves for editing geometry processing. Despite recent progress the area, automatic tools often produce spurious branches or incorrect connectivity around curve junctions; smooth out sharp corners. These issues detract from use such tools, both an aesthetic viewpoint feasibility downstream applications (e.g., coloring inbetweening). We address these problems by introducing novel line algorithm that splits into three components: (1) finding keypoints, i.e., endpoints, junctions, corners; (2) extracting topology, connections between keypoints; (3) computing those connections. compute optimal connecting via geometric flow --- PolyVector Flow aligns drawing, disambiguating directions Y-, X-, T-junctions. show our system robustly infers topology detailed complex drawings. validate quantitatively qualitatively, demonstrating method visually outperforms previous work.