Multi-pose face synthesis from a single 2D reference image

Face image synthesis under different poses from a single 2D reference image is challenging. Since self-occlusion exists in profile face images, the unknown information of the reference face in other poses has to be synthesized by the prior knowledge. In this manuscript, we separate facial appearance into texture and shape information with factor-specific active appearance model. We aim at revealing the sparse structure of facial texture patches under different identities and the latent continuous facial geometric shape manifold of the pose, which can be used as the general prior information to guide the facial texture synthesis under different poses. To obtain a general pose manifold, tensor analysis is applied on the shapes to acquire the low-dimensional and general pose information. The facial texture is divided into homogeneous patches by Delaunay Triangulation strategy according to the facial geometry defined by the pose manifold. The seen texture of the synthesized poses is transformed from the reference image. And the occluded local texture is synthesized from its sparse neighbors. The experiments show the proposed method can synthesize high quality multi-pose faces, which can be recognized more easily.