3D Face Model Reconstruction From Video

• Main Authors: Chia-ming Cheng, 鄭嘉民
• Other Authors: Shang-Hong Lai
• Format: Others
• Language: en_US
• Published: 2001
• Online Access: http://ndltd.ncl.edu.tw/handle/02884898102911166922

碩士 === 國立清華大學 === 資訊工程學系 === 89 === In this thesis, a complete system is presented for 3D face model reconstruction. We focus on how to get a 3D face model from an input image sequence. The proposed method is based on adapting a generic 3D face model to the 3D locations of some feature points recovered from an image sequence. The techniques involved in our system include face pose estimation, structure from motion, structure from silhouette, model adaptation, and texture mapping. First of all, the corresponding vertices of some predefined features on the 3D generic and images are extracted. Subsequently, an approximate face pose for each image frame is estimated by using an iterative pose estimation method. Then we present a robust bundle adjustment algorithm to recover the 3D structure of a set of feature points and refine the face pose estimation. Based on the estimated pose information, the structure from silhouette algorithm is proposed to further improve the 3D model. Combining the 3D geometric constraints recovered from the bundle adjustment and structure from silhouette, we deform the 3D generic face model by using a radial basis function interpolation to fit these geometric constraints, thus reconstructing a personalized 3D face model. Finally, the texture map of the model is computed by integrating multiple images with an appropriate weighting scheme. Several textured face models from image sequences of males, females and children are shown.