Implementation of Everybody Dance Now by Chan, et al.
Intro • Key Features • How To Use • Download • Related • Requirements •
This is an implementation of a research paper, Everybody Dance Now. The main objective of the project is to allow frames of poses to be synthesized. This in turn can be used for showcasing dance moves or any movement. The paper utilizes pix2pixHD generative adversarial model to synthesize an image from semantic pose heat maps. The poses are obtain from openPose framework. Heat maps and part affinity maps are used to perform pose estimation. The pix2pixHD GAN introduces, several separate inputs to the generator and the discriminator, while having multiple decriminators that advance the photo realistic depiction of the synthesized image.Further details about the pix2pixHD GAN can read on NVIDIA's github or in my blog.
Create gifs or videos of synthesized dance moves
Make people you know dance!
Simple use of GANs and CNNs
Clone or download the repo
Obtain target dataset - the person you want to perform the generated poses on.
Obtain the test dataset to source the pose estimation from
Train the GAN model on the target dataset
Transfer the models learnt outcome to output synthesise images, using generator and the semantic label heat map of the poses.
Create video from moving frames.
Clone the github repo including source.
LinkedIn Damini K Rijhwani
GitHub @DaminiR
AI Blog Disrupt Culture