Today I’ve got a dance transfer NN that can make anyone a good dancer. All you need is a source video of a dancer and 20 mins of you moving in front of a camera.
The smoothness of the shown GAN (compared to classical Pix2pix approach) is achieved by the Generator looking at the current pose frame and the last generated frame. The Discriminator also works with pairs of poses and images.
Full paper: https://arxiv.org/pdf/1808.07371.pdf