Spatio-temporal Caricatures of Facial Motion

It is well established that there is a recognition advantage for slightly caricatured versions of static pictures of faces (e.g., Rhodes et al., 1987, Cognitive Psychology, 473- 497; Benson Perrett, 1994, Perception, 75-93). Recently, similar caricature eects have been shown using temporal or spatial exaggerations of complex body movements (point light displays) (Hill Pollick, 2000, Psychological Science, 223-228; Pollick et al. 2001, Perception, 323-338). Here, we generated spatio-temporal caricatures of facial movements using a motion morphing technique developed by Giese Poggio (2000, International Journal of Computer Vision, 59-732000) to investigate whether identication from facial motion can be improved by caricaturing. The motion caricaturing was accomplished using hierarchical spatio-temporal morphable models (HSTMM). This technique represents complex motion sequences by linear combinations of learned prototypical movement elements. Facial motion trajectories of 72 re ecting markers were obtained using a commercial 3D motion capture system (VICON). These original trajectories and the morphed or exaggerated versions are applied to photo-realistic head models (Blanz Vetter, 1999, SIGGRAPH: 187-194) using a commercial face animation software (famous3D Pty. Ltd.). In a rst experiment which employed motion data captured from 2D videos, we tested the quality of this linear combination technique. Naturalness ratings from 7 observers were obtained. They had to rate an averageshaped head model, which was animated with three classes of motion trajectories: 1) original motion capture data, 2) approximations of the trajectories by the linear combination model, and 3) morphs between facial movement sequences of two dierent individuals. We found that the approximations were perceived as natural as the originals. Unexpectedly, the morphs were perceived as even more natural (t(6)=4.6, p<.01) than the original trajectories and their approximations. This might re ect the fact that the morphs tend to average out extreme movements. In a second experiment 14 observers had to distinguish between characteristic facial movements of two individuals applied to a face with average shape. The movements were presented with three different caricature levels (100, 125, 150). We found a signicant caricature eect: 150 caricatures were recognized better than the non-caricatured patterns (t(13)=2.5, p<.05). This result suggests that spatio-temporal exaggeration improves the recognition of identity from facial movements. We are currently investigating whether this result generalizes to the 3D motion data and to dierent types of facial motion (e.g., rigid head motion versus non-rigid deformation of the face)