Recurrent convolutional adversarial networks for generative modeling of human motion

Abstract

In recent years, the importance of Generative Adversarial Networks (GANs) has been discovered and their potential uses in the near future became obvious. In this thesis, the idea of a machine learning model based mainly on Generative Adversarial Networks will be explained in detail along with the tools used to complete it A Generative Adversarial Network is a machine learning model that consists of two basic parameters. The first parameter is G (Generator) and the second is D (Discriminator). After training, the model is capable of doing two things that are defined by its parameters. First, it can produce samples through G, which are related to the samples which the model was trained with, the training data. On the other hand, it is also able to discriminate if a sample was produced either from G or from the training data through D. The model referred to in this dissertation combines GANs along with Convolution techniques for processing input data for training. In addition, a combination of Recurrent Neural Networks (RNNs) and more specifically the fitting of a Long Short Term Memory (LSTM) has been made for a better convergence of the algorithm due to its ability to memorize features and extract dependencies between human movements. Finally, this model is able to produce video samples that present human motion, but also to separate whether if these samples came from G or from the training data.