A novel frontal view synthesis method based on neighbor embedding

This paper presents a novel approach that can efficiently synthesize a virtual frontal view, given only a single non-frontal face image. A non-frontal face image is separated into shape and shape-free texture, and Neighbor Embedding (NE) is applied to them respectively. The virtual frontal face can be generated by warping the shape-free texture to the shape and enforcing local compatibility and smoothness constraints between adjacent patches. While our method resembles other learning-based methods in relying on a training set, our method is novel in that it accurately reveals the intrinsic distribution of different pose feature spaces by assuming that the feature spaces for the frontal and non-frontal face images share similar local manifold structure. Experimental results show that the proposed method is better than Linear Object Classes (LOC) based method and Tensor-based Subspace Learning (TSL) method, both in the subjective and objective.