Facial modeling and animation are valuable for both application and research field such as animation film, human-computer interface, medical treatment and education. In this thesis, we focus on some of the key issues in image based automatical modeling and animation of individual face. The main contributions include: 1.We propose a reconstruction method under the control of key points, which can generate individual face model fast. Based on one frontal face image captured from PC camera, the facial key points can be detected automatically and their 3D coordinates are estimated with morphable models. Then a generic face model is deformed with RBF and key points' coordinates to generate individual mesh. Finally, we crop the face region from input image and generate individual texture to make the model more realistic. The method can be applied in multi-player online games. 2.For one face image captured in different direction and different illumination, we propose a totally automatical method for reconstruction. We only need the result of face detection to initialize parameters for individual face. The final parameters include morphable models can be obtained by matching rendered model image to input image. The regular mesh facilitates normalization of input samples and generation of individual face. The strategy of iteration by stages makes computation faster, also it can eliminate errors when the background is similar to face region. 3.We improve the automatical reconstruction method by replacing rendered image with shape-free texture. Thus we can eliminate the influence of posture and shape variance, and can decrease the probability of local convergence. Also we propose visual angle map in the method. It improves the key points alignment and refines detail of texture image. 4.We introduce spherical parametrization in facial animation and achieve animation remapping with sparse key points. The original animations are captured by Motion Capture system and with only 34 markers. After remapping the key points' motion vectors, we interpolate the non-key points' motion vectors with barycentric coordinates of key mesh triangles. All the method can achieve fast remapping of realistic expression animation. Fewer markers makes fewer uncomfortableness for the performer.
修改评论