| dc.description.abstract | Stylized characters are highly used in movies and games. Furthermore, stylization is mostly preferred over realism for the design of toys and social robots. However, the design process remains highly subjective because the influence of possible design choices on character perception is not well understood. Investigating the high-dimensional space of character stylization by means of perception experiments is difficult because creating and animating compelling characters of different stylization levels remains a challenging task. In this context, computer graphics algorithms enable the creation of highly controllable stimuli, simplifying examination of specific features that can strongly influence the overall perception of a character.
This thesis is separated into two parts. First, a pipeline is presented for creating virtual doubles of real people. In addition, algorithms are described suitable for the transfer of surface properties and animation between faces of different stylization levels.
With ElastiFace, a simple and versatile method is introduced for establishing dense correspondences between textured face models. The method extends non-rigid registration techniques to allow for strongly varying input geometries. The technical part closes with an algorithm that addresses the problem of animation transfer between faces. Such facial retargeting frameworks consist of a pre-processing step, where blendshapes are transferred from one face to another. By exploring the similarities between an expressive training sequence of an actor and the blendshapes of a facial rig to be animated, the accuracy of transferring the blendshapes to actor's proportions is highly improved. Consequently, this step overall enhances the reliability and quality of facial retargeting.
The second part covers two different perception studies with stimuli created by using the previously described pipeline and algorithms. Results of both studies improve the understanding of the crucial factors for creating appealing characters across different stylization levels. The first study analyzes the most influential factors that define a character's appearance by using rating scales in four different perceptual experiments. In particular, it focuses on shape and material but considers as well shading, lighting and albedo. The study reveals that shape is the dominant factor when rating expression intensity and realism, while material is crucial for appeal. Furthermore, the results show that realism alone is a bad predictor for appeal, eeriness, or attractiveness. The second study investigates how various degrees of stylization are processed by the brain using event-related potentials (ERPs). Specifically, it focuses on the N170, early posterior negativity (EPN), and late positive potential (LPP) event-related components. The face-specific N170 shows a u-shaped modulation, with stronger reactions towards both, most abstract and most realistic compared to medium-stylized faces. In addition, LPP increases linearly with face realism, reflecting activity increase in the visual and parietal cortex for more realistic faces. Results reveal differential effects of face stylization on distinct face processing stages and suggest a perceptual basis to the uncanny valley hypothesis. | en_US |
