Human Moving 3D Pose Generation Using Conditional Variational Autoencoder

In recent years generative models have been a viral topic within the field of deep learning. First, generate model used to generate the picture and then temporal feature comes into this new area of research and many researches have been done to generate pixel-based video. Alongside generative models, human action recognition is a well studied and reliable task in the deep learning field, the goal of this task is to determine the type of action of human activity based on the sequence of moving skeleton by using the spatiotemporal features, but in this work, we tried to do this task reversely, generate moving human skeleton by choosing action type. Although synthetic human motion without a deep learning approach exists, mostly they are not realistic enough. So in this work, we defined our goal to generate realistic and plausible human motion sequences in 3D, and we believe that the deep learning approach can solve this problem. In this work, we used an unsupervised generative model to generate a sequence of moving human skeletons based on a specific action. The Conditional Variational Autoencoder(CVAE) model has been used to create conditioned latent space. By sampling from the conditioned latent space we can generate desired human motion based on the action. To continue we used and evaluated three different layers as the backbone of CVAE, which are Gate Recurrent Unit(GRU), Long-short Term Memory(LSTM), and Graph Convolutional Layer(GCL). We trained the three CVAE models with the same parameters and dataset and in the end, we compared and evaluating the efficiency of CVAE models with each other by using four different metrics.