Overview

Machine learning-based approaches require a large amount of training data to perform well on previously unseen test data. If there is no available real annotated training dataset a synthetic one can be rendered and used. Regarding all human body-related tasks, it is important to make the virtual data qualitatively as similar as possible to the real human body scans, so that the statistical models are able to generalize well in real-world scenarios. Therefore, we would like to enhance SMPL-based models with cloth, hair, or other details missing in the SMPL-X models and render the synthetic data as similar to real scans as possible. 

Goals

The main goal of the thesis is to augment synthetically generated data to adapt the synthetic domain to the domain of real human body data. Add more details to the generated human body data, such as clothes, hair, and realistic background scene. Moreover, the aim is to evaluate the data by using it as a supervision signal to train and test a machine learning model. 

Roadmap