Realistic face modeling with robust correspondences

Finding robust correspondence is an important problem in structure from motion algorithm. Because the human face contains many low texture and homogeneous areas, some algorithms such as corner matching are unstable and may fail sometimes. We used the face definition parameters and the symmetry of human face as prior knowledge to find reliable correspondences between two pictures, while most SFM algorithms use the generic model as a modulator in the post-processing steps. This work proposes a whole scheme to construct textured 3D face models from two views with a few user interactions. According to the correspondences, a multistage SFM approach is used to reconstruct the structure. Then we use the RBFCS algorithm to interpolate more 3D points according to the scattered feature points. A user with an ordinary camera can use our system to generate his face model in a personal computer.