| Creating realistic human characters is one of the most challenging topics in computer graphics, due to the subtleties of human motion and appearance, and how sensitive we are to the appearance of our own kind. In this thesis, we will discuss methods for creating realistic, animatable models of human body surface shape, based on data acquired from the real world. We will present several algorithms for taking unstructured shape data, such as that obtained from a 3D laser range scanner, and fitting a human character model to the data.; We will focus on three main problems in human shape modeling. First, using a large corpus of individual scans, we will discuss how to analyze the data to model the individual variation in body shape between people of different body types. We present several interesting and novel synthesis methods that make the problem of creating and editing digital characters much simpler.; Second, we will examine how an individual's body changes shape as it changes pose. Rather than explicitly modeling the underlying phenomena of bulging muscles and folding flesh, we will learn these pose-dependent deformations as arbitrary functions whose parameters can be learned from data.; Finally, we will consider the pose-dependent and individual shape variation problems in tandem, thereby creating a shape model that can generate any body shape in any pose. Our approach is very general, and can tolerate irregular samplings in pose and individual body space. |
