Generating synthetic images of human skeletal motion for pose and kinematics estimation tasks

We developed a software that generates images of human motion from kinematics calculated using musculoskeletal modelling. The images are automatically annotated with information from the underlying skeletal model, including 3D positions of joint centers. The software enables the generation of an arbitrary number of images from a small number of skeletal poses by varying visual factors such as camera angle, background, body morphology, and skin and clothing textures of the person. The generation of synthetic images can be helpful in generating training data for supervised learning-based human pose estimation and motion tracking models. Because our software uses information from biomechanical models of the human musculoskeletal system, its annotations have the potential to be more accurate than those of existing large datasets of real images, where non-experts have marked the positions of anatomical landmarks. Additionally, new annotation points can be defined by editing the virtual marker set of the musculoskeletal model, which allows the generation of images with user-defined annotations.