Multi-view face image synthesis using factorization model

We present a sample-based method for synthesizing face images in a wide range of view. Here the "human identity" and "head pose" are regarded as two influence factors of face appearance and a factorization model is used to learn their interaction with a face database. Our method extends original bilinear factorization model to nonlinear case so that global optimum solution can be found in solving "translation" task. Thus, some view of a new person's face image is able to be "translated" into other views. Experimental results show that the synthesized faces are quite similar to the ground-truth. The proposed method can be applied to a broad area of human computer interaction, such as face recognition across view or face synthesis in virtual reality.