Learning style translation for the lines of a drawing

An important task of pattern recognition is identifying or synthesizing stylistic variations on a signal independently of its underlying content [Omohundro 1995; Hofstadter 1995; Tenenbaum and Freeman 2000; Brand and Hertzmann 2000; Rose et al. 1998]. For example, one may want to differentiate stylistic variations in handwriting from the underlying content, or letters. We study this problem for the particular task of rendering line drawings. We want to allow the user of a drawing program to apply a new style to selected lines of the drawing, while keeping the “content”, or the overall shape of those lines, constant. Present systems typically provide such style control through parametric or interactive manipulations. For example, the thickness of the line can be adjusted, or the Fourier spectrum of the x-y coordinates of the line can be modified to yield a different line character ([Finkelstein and Salesin 1994; Unama et al. 1995; Bruderlin and Williams 1995; Hsu and Lee 1994]). User guidance is required in the rendering or font appearance models of [Winkenbach and Salesin 1994; Zongker et al. 2000]. 3-d object properties are used for the line-rendering methods of [Hamel and Strothotte 1999]. But even advanced parametric manipulations represent a small subset of the possible stylistic variations that a designer might consider. Figure 1 shows a set of line strokes each made in a range of different line styles. This range of styles would be very difficult to describe parametrically. Yet we would like to build a system which supports content-preserving translation across such different styles of lines. See, for example, [Borgman 1977], for examples of lines drawn in different styles. We propose a learning-based approach to this problem. The system designer collects examples of many different lines, each drawn in the various styles that the system is to contain (the “training data”). An artist, not a programmer, decides what it means to render a particular line over a range of styles. The system refers to this training data in order to automatically translate lines made by a user into a particular desired style. Our system has similarities to the interpolation-based style-modification system of [Rose et al. 1998]. Their “adverbs and verbs” correspond to what we call “style and content”. However, we emphasize automatic operation at runtime; given new lines, the system must both identify the corresponding content (training line), as well as translate to a novel style. Our example-based approach in the domain of line modification is analogous to examplebased texture transfer methods [Hertzmann et al. 2001; Efros and Freeman 2001].