Lowering the Cost of Virtual Human Rendering With Structured Animated Impostors Amaury Aubel, Ronan Boulic and Daniel Thalmann Computer Graphics Lab Swiss Federal Institute of Technology (EPFL) CH 1015 Lausanne, Switzerland Tel: +41 21 693 52 14 Fax: +41 21 693 53 28 e-mail: {aubel, boulic, thalmann}@lig.di.epfl.ch ABSTRACT We present in this paper an image-based rendering technique to represent virtual humans in real-time. Though we deal only with virtual humans, the concept can be applied to any articulated character. Our technique is termed animated impostors because it extends the notion of dynamic impostors to take into account changes in the character’s appearance. We also show how impostors can be embedded into large-scale simulations/applications. Keywords: virtual humans, impostors, LOD, sprites, image-based rendering. 1. INTRODUCTION A lot of research has been devoted over the past decades to model and animate virtual humans in a realistic fashion. As a result, today’s animators have tools that enable them to create so compelling synthetic actors that our visual system is sometimes mistaken, as in the recent film “Titanic”. But realistic synthetic characters are not confined to the sole movie industry. They are being used increasingly in real-time applications such as computer games, collaborative virtual environments and for digital communities. However, one of the current limitations to displaying numerous virtual humans lies in the number of polygons that can be processed by the rendering pipeline. Geometric level of detail (LOD) is the most common answer to this problem. Pratt et al. [Pratt97] applied this technique to virtual humans in large-scale, networked virtual environments. In their application each simulated human has four different resolutions ranging from about 500 polygons down to only three polygons. This may be satisfying from a simulation point of view, for the system manages to sustain a high refreshment rate. In counterpart, the result is necessarily poor in terms of realism and aesthetics because there is no way a virtual human made up of a few dozen polygons can look good, no matter the viewing distance and the designer’s skills. Our approach to rendering virtual humans is based on the fact that there is a high degree of coherence between two successive frames. Even when rendering a complex, moving, self-deforming figure such as a virtual human, the changes from frame to frame in the animation remain small. Besides, since the human brain tends to reconstruct the missing data when shown only certain frames, a few key