Semi-Supervised Pipeline for Human Sprites Neural Rendering

Abstract

In this paper we are discussing the possibility of a hybrid approach in renderable scenes. The main idea of the presented experiment is to render the human actors by using existing videos of the characters. The input video is first converted to a sprite dataset. The dataset is generated with supervised techniques but human intervention is also required. After that we extract body and pose parameters. Lastly, we render novel poses using a GAN-based approach similar to pix2pix.