A face wrapping method based on pose-specific shape eigenspace

Abstract

Generating virtual face images with different poses has potential applications in many areas, such as face recognition, human-machine interaction, portrait combination, and computer graphics. However, in some situation, the available face images are quite limited, which makes the problem difficult. This paper proposes a pose-specific shape eigenspace based face wrapping method to generate virtual face images with different poses from a specific pose. A predefined training set is necessary. According to their poses, training faces with annotated landmarks are manually divided into several groups, each of which is utilized to learn a pose-specific shape eigenspace by K-L transform. For a new image under a certain pose, its shape information described by the annotated landmarks is firstly projected to the expected pose-specific shape eigenspace to represent the shape information of this image under the expected pose. Then, all corresponding points between the represented shape and original shape the are matched and the texture information of all points in the represented shape are covered by the gray or color information of the corresponding points in the original image to generate a virtual face image under expected pose. To quantify the similarity between the generated virtual images and real images, cosine similarity is adopted. Experiments on IMM, PIE and YaleB face subsets show that the similarity of the virtual image and real images is over 0.9, no matter there is high or low similarity between test set and training set, which illustrates the effectiveness of the proposed method.