Representation of material properties has a long history in the graphics community. Schlick [14] provides a good survey of this literature. Our paper takes a different approach, focusing on reproducing shading algorithms used by an artist in their work. In NPR, only the areas of cartoon shading [10] and technical illustration [5] have incorporated color shading techniques. There have been several researchers outside of the NPR community who have tried to capture lighting effects from one medium and apply them to another scenario. Work by Miller and Hoffman [12], Debevec et al. [3,4], and Sato et al. [13] captured lighting effects in order to plausibly embed computer graphics objects in photographs or video or to create new scenes under the same environmental conditions. Reverse engineering of BRDFs was applied to the domain of photorealism by Yu et al. [16,17]. In this paper, we attempt to solve a similar inverse problem to reproduce artistic shading by sampling art work. One notable departure from prior work is that no attempt is made to separate lighting information from texture. This is due to the incomplete information available in a single artwork sample. Our techniques are grounded in the ideas for environment maps presented by Blinn [2] and Greene [7], with slight variations on the mapping to the target surface. The lit sphere approach is similar in spirit to results obtained by Heidrich and Seidel [8].
To understand the proposed method for extracting artistic shading models, it is useful to review techniques utilized in the rendering community for capturing reflectance. Suppose we have a small patch of a given material within an environment, and we wish to measure how much light leaves the surface with respect to a set of outgoing directions. We can accomplish this by keeping the environment and measuring device fixed and rotating the patch about its center. This procedure allows us to determine a correspondence between the direction of the surface normal and the amount of light which reaches the device.
A sphere provides coverage of the complete set of unit normals. Assuming the sphere is small with respect to the scale of the scene, the light arriving at all nearby points on the sphere will be roughly the same, allowing us again to relate surface normal to reflected light for a particular material. Thus, approximate measurements can be obtained by photographing a sphere made of the given material embedded in the identical scene whose radius is small compared to the scale of the scene (see Figure 2).
Finally, suppose we have an object of the same material as the sphere, and similar scale. Under these conditions the reflectance of the surface and the character of the incoming light will not vary much from point to point on the surface. If this is true, then we are justified in replacing the sphere with the desired surface and lighting it according to surface normal using the data from the photographed sphere. In essence, the sphere serves as a surrogate for more complex objects in order to simplify the characterization of reflected light and assure sufficient coverage of normals. We refer to this ``paint by normals'' method of shading as the lit sphere model. Because our model is informed only by light which leaves a surface in the direction of the eye, it will fail to distinguish between variations due to lighting and those due to texture. Therefore, our method will allow the user to render scenes from novel viewpoints, but will exhibit artifacts under animation as the texture swims across the surface to follow the eye. We shall explore the ramifications further in Section 5.
