Realistic Facial Expression Synthesis

Facial expression animation is an important branch of computer graphics, which is always the research hotspot. People have generated plenty of results in facial expression, and many of them have been applied in film and television, advertisement, and games, etc. The famous movies, such as King Kong, The Lord of the Rings, and Avatar used facial expression synthesized with computer successfully, which demonstrated the ever-lasting charm of facial expression animation, and gave people deep impression about this research domain. With the development of technology and the progress of the times, people require more realistic results, and hope the calculation can be faster. Broad application prospects and the technical feasibility must lead to more and more focus and investment.This thesis described the related work of facial expression synthesis, category the existed methods into several sets, and analyzed the advantage and disadvantage of each set. Based on this, we focused on the research of several key problems in realistic facial expression synthesis, and provided some systematic resolutions, including the acquisition of facial motion data and the extraction of facial expression, the synthesis of realistic facial expression, and the editing of facial animation, etc. To be specific, our work includes the following aspects:In this thesis, we presented a scheme to extract high fidelity facial expression. For the facial motion data captured from a Vicon motion capture system, we mapped it to the space of the neutral facial model using Radial Basis Functions (RBF) interpolation. With the aid of the neutral markers during motion capture, we extract the facial expression of the performer, and obtain the corresponding rigid head motion accordingly.This thesis presented an expression synthesis method based on Laplacian deformation, which can preserve the model details during facial deformation and improve the reality of the synthesized expression. For the given facial model, we calculate the Laplacian coordinates of all the vertices firstly. During the expression synthesis, we preserve the vertices’ Laplacian coordinates. With the displacements of the facial feature points and the positions of the fixed points, all the new positions of the facial model are evaluated, which correspond to a new facial expression. With the rigid head motion, we synthesized the target facial model with similar expression and pose corresponding to the performer.This thesis proposed to synthesize facial expression via geodesic distance and RBF interpolation. For a mesh surface, the Euclidean distance may be quite different from the distance along the mesh surface, and this is more obvious for a facial model with holes, such as the mouth and the eyes. The geodesic distance provides an approximate description for the distance along the mesh surface, which is important for expression synthesis. We present a new scheme to synthesize natural facial expression using the geodesic distance instead of the Euclidean distance in the RBF interpolation.To improve the post-editing efficiency of the synthesized facial animation, this thesis presented a space-time-based method to edit the facial animation. It propagated the users’editing among the input facial model with the Laplacian-based mesh deformation. Then, a Gaussian function was employed to propagate the editing effect to the surrounding frames in a fast way. It provides users with the local control of the facial animation editing. Users can assign the influence range of their editing and the editing result is propagated to the surrounding frames naturally.This thesis proposed to edit the facial expression based on2D shape deformation. Expression editing is performed to minimize the distortion of all the triangles of the facial model. Taking into the consideration of the face properties, we try to preserve the total length of the facial model’s silhouette during the deformation.In the meantime, the technology of shape deformation can also be used to garments design to calculate the garment in new poses automatically. It can help designers efficiently present their ideas in a professional manner, detect any inappropriateness in the designs, provide instant feedbacks on design and thus accelerate the design process. We use the human skeletons to drive the garment deformation, and the handles and target positions in the new pose are obtained accordingly.We performed the above methods with several facial models, and obtained satisfactory results for all. At last, we conclude our work, analyze the existed problems, and indicate the future work.