Generative adversarial networks shows amazing results in list of tasks where we want from neural network to model our data distribution and produce realistic outputs, based on some condition. We will analyse recent state of the art approaches for conditional image synthesis trying to solve problem of human face and pose transferring. Main idea is to be able to generate realistic images of person B moving in a pose of any other person A. So, the task consists of 3 stages:
- human pose recognition of person A (pose keypoints extraction),
- face landmarks detection of person A
- image synthesis of person B, conditioned on face and pose keypoints
On a way of solving this tasks we will face interesting problems of finding right methods to encode input labels, labels denoising, and improving synthesized image resolution while saving GPU memory on training stage.
Also, we will review techniques of conditional GAN trainig stabilization and analyse problem of frame consistency for frame-by-frame pose transferring using GANs.
For the last five years, Oles has been working as a machine learning engineer at VideoGorillas. He develops computer vision applications for film studios, video stream object detection systems, and applied NLP systems including neural network-based language modeling. Previously, he built automated analytics systems for social media. At ODSC, Oles will present some recent research results in the field of Generative Adversarial Networks done in cooperation with the NeoCortext team.
Event Timeslots (1)
Track B (Lower Floor)