Artificial Synthesis of Single Person Videos through Motion Transfer using Cycle Generative Adversarial Networks and Machine Learning

In this work, we address the problem of motion transfer from the source video of one person to another target person’s video. This basically involves human pose synthesis, whereby the source person’s poses in a video are superimposed frame-by-frame onto the target. This results into an artificially generated video of the target, where his/her movement is determined by the given source video. In this work, we use global and local edge detectors as well as pose or skeletal keypoints ex- tractors for synthesizing an artificial human dance video. In this paper, we explore two different Generative Adversarial Network (GAN) architectures for pose detection and superimposition from source to target object. We present experiments and results to validate the proposed models’ outputs, as well as compare their performances.