Geometric Modeling for 3D Human Pose Estimation and Motion Transfer

This thesis explores the geometric modeling to perform 3D human pose estimation and human motion transfer. We first propose the scaled orthographic projection to alleviate the need of ground-truth 3D poses for in-the-wild images. Then, we utilize the fisheye camera model to solve the distortion problems caused by fisheye cameras from first-person and third-person views for 3D human pose estimation. To verify the effectiveness of the proposed method on real-world images, we collect a dataset for multi-person 3D human pose estimation from fisheye cameras. Finally, we exploit the relation of 2D and 3D information to perform the human motion transfer with pose consistency. Findings of this thesis may lead to development of the more advanced computer vision algorithms focusing on 3D object reconstruction from fisheye cameras and motion transfer.