This is published under the terms of CC BY-NC.

Downloaded: 1915 times

Abstract

Deepfake technology has been undoubtedly growing at a rapid pace since 2017. Particularly since using GAN architecture was popularized, research in this area has grown and seems to only be gaining momentum. One interesting area is animating images of full body humans using deep learning. This paper looks at the research done in this area and research that can influence it by looking at papers regarding human pose transfer, human motion transfer, and human motion generation. All of these types of papers have similar requirements, where a target pose must be abstracted to a skeleton and combined with appearance data from a source image to generate a result. The primary difference in the three types of research is whether or not there is motion in the result and whether that motion is given as an input or generated by the model. Overall, the research in this area is still new, and with the potential applications of this technology, both good and bad, there are many avenues of potential future research in this area in both creation and detection.