Abstract
To realize natural human interface, there are many studies to synthesize facial expressions using computer graphics. In this paper, our goal is to reconstruct a 3D structure of a face with arbitrary expressions (expression-B) using 3D structure with expression-A, two images with expression-A, and domain knowledge of expressions. First, from input data, two projection matrices, which project the 3D structure to two input images, are estimated. Second, deformation from expression-A to expression-B is calculated as the difference vector of corresponding points in the images with expression-A and-B, where the images with expression-A and-B are from the image sample database. By using the difference vectors, similar property of face, symmetrical property of face and geometrical constraint, the images with expression-B are estimated. It two projection matrices and two facial images with expression-B are given, 3D structure with expression-B can be reconstructed. Experimental results show that 3D structure with expression could be reconstructed with fairly good performance.
