Nonlinear Synthesis of Expression Variation Dynamics on Video Using Deep Dynamic Bottleneck Neural Networks
Saeed Montazeri, Seyyed Ali Seyyedsalehi, Nima Amini · 2017
Nonlinear components extracted from a deep structure of bottleneck neural network have a great ability in expressing input space in a low-dimensional manifold. Sharing and combination of components makes it possible to synthesize and interpolate a new data. Presented paper introduces a novel Dynamic Deep Bottleneck Neural Network to analyze videos of emotion expressed on the faces of input subjects. The main extracted features are identity, emotion and expression intensity that are lied in three different subspaces of a manifold. This model shows potential to synthesize new videos showing variations of one specific emotion on the face of unknown subjects. Produced videos show variations from neutral to apex of an emotion on the face of unfamiliar test subjects which is in average 80% similar to the reference videos in the scale of SSIM method.