Learning orthogonal factorization in gan latent space

Abstract

A method for learning disentangled representations of videos is presented. The method includes feeding each frame of video data into an encoder to produce a sequence of visual features, passing the sequence of visual features through a deep convolutional network to obtain a posterior of a dynamic latent variable and a posterior of a static latent variable, sampling static and dynamic representations from the posterior of the static latent variable and the posterior of the dynamic latent variable, respectively, concatenating the static and dynamic representations to be fed into a decoder to generate reconstructed sequences, and applying three regularizers to the dynamic and static latent variables to trigger representation disentanglement. To facilitate the disentangled sequential representation learning, orthogonal factorization in generative adversarial network (GAN) latent space is leveraged to pre-train a generator as a decoder in the method.