Human Imitation in Images and Videos using GANs

The paper aims to generate a pseudo-Image/Video from a given input Image and a reference video where the features of the human in the reference video are transferred to the human in the input Image. The source features, such as texture, style, color, and face identity, can be preserved using a Liquid Warping GAN. A three dimensional body mesh of the subject can be given to the GAN so that it propagates the source features in both image and feature spaces, and synthesizes an image with respect to the reference image or video. Full body motion imitation is a challenge in Artificial Intelligence and requires several modules in order to perform effectively and efficiently. To combine to different sources of appearance and motion into a single cohesive video joining them both is both interesting and useful for many applications. iPER dataset is used to train the GAN. The dataset is taken from Shanghai Vision and Intelligence Perception Lab. Within the dataset, there are different conditions of shape, height and gender dispersed among 30 human subjects. Each human subject wears separate clothes and performs a video with random actions. This technique can be employed for animations where an animator doesn’t have to draw and render multiple sequences where an actor can just perform the sequence physically and this can be applied to multiple humans, thus reducing time and effort. This method can also be used to transfer the style of clothing which has several applications while shopping online.