The original advantage of stippling was its ease of reproduction. The half-toning used to print images in books was of highly variable quality and often drawings were drastically resized to meet space requirements. While normal drawings suffered from such treatment stipple drawings retained their attributes more faithfully. In addition, printing a stippled drawing requires only the ability to produce dots of a single colour, making it an inexpensive technique .
However, stippling has significant artistic merit independent of its utility. The stipples can represent fine detail and texture with little cost in complexity. Stippling is particularly good at clearly representing smooth, rounded objects without sharp edges and so is often used in medical and archaeological texts.
We wish to generate stipple drawings from images with as little user input as possible. The goal is to develop a tool which can generate high-quality stipple drawings from any source whatsoever, which implies that we use images as input and not 3D models. While this limits the amount of information we have to work with, it allows us a greater variety of input sources. For example, a user could start from a scanned pencil sketch, a photograph, the output of a 3D interactive application, frames of an animation, etc.
One of the features of a good stipple drawing is that the stipples are well-spaced, that is, the stipples do not clump together, leave uneven voids, or form unwanted patterns. The artist achieves this by carefully placing each stipple onto the page, explaining why stipple drawings often take weeks to create by hand.
Central to our approach is the use of centroidal Voronoi diagrams to produce good distributions of points, as explained in Section 2.1. These distributions can be pre-computed for various different constant tonal values and accessed at run-time to generate stipple drawings rapidly, as covered in Section 4. Alternatively, the input image can be used directly as a weighting function to create a distribution of points that approximate its tones. This method produces images of higher quality but takes more processing time, as explained in Section 3.
Hausner  uses an approach similar to our iterative technique outlined in Section 3 for aligning the rectangular tiles of a decorative mosaic. Their approach differs significantly from ours in that they must align the tiles' orientation in addition to their position. However, edges that are to be preserved by the algorithm must be entered separately by the user. This makes the algorithm less suitable for a non-interactive application.
Adrian Secord 2001-11-20