An Efficient 3D Synthetic Model Generation Pipeline for Human Pose Data Augmentation
3D modeling of articulated bodies of humans or animals and using these models for synthetic 2D and 3D pose data generation can mitigate the small data challenges faced by many critical applications such as healthcare. In this paper, we present our efficient 3D synthetic model generation (3D-SMG) pipeline used for body pose data augmentation. 3D-SMG pipeline starts with scanning point clouds from various angles around the subject using an off-the-shelf RGBD camera. We then implement a dual objective iterative closest point (ICP) algorithm that uses both color (if available) as well as geometric information from point cloud and apply a pose graph node optimization to form one single rigid body mesh. 3D-SMG also includes a series of post processing steps to obtain a smooth mesh at the end of the pipeline. The approach allows it to be applied to any articulated object such as a human body or an animal. Our experiments also show high level of accuracy in dimensions of obtained 3D meshes, when compared to the original subject. As the final step towards developing augmented pose dataset, we perform model rigging to articulate the 3D model of the subject and generate dynamic avatars within variety of context-feasible poses.