More optimal strokes for NPR sketching

Sketching is a drawing style where approximations and successive refinement in the drawing process are evident. The approximation of contours in sketching involves multiple overlapping strokes that are relatively long in regions of low curvature and shorter in high-curvature areas, yet unimportant high-curvature details are omitted in the initial stages of a sketch. Rendering contours with a single long stroke does not capture the feel of a sketch, and a simple strategy of breaking strokes at curvature maxima is easily confused by unimportant details and noise. We address the contour breaking problem for sketching by clustering samples of the contour based on proximity and orientation, making use of a global clustering algorithm (normalized cuts). The strokes generated by this approach qualitatively resemble those produced by real artists, and the successive approximation effect seen in sketching can be simulated by employing our approach at a succession of scales (increasing the number of clusters).