Breaking the 2D Barrier: 3D Human Pose Reconstruction with Graph Neural Networks

Reconstruction of 3D human poses has been a key area of research with applications in various fields, such as animation, healthcare, and sports analysis. It typically consists of two steps. First, by detecting landmark points from an image or video, one estimates 2D poses. Then, “lifting” these estimated 2D poses into a 3D model representing the human form is performed. In this paper, we propose a pipeline for 3D human pose reconstruction that converts 2D keypoints into 3D coordinates using PoseGraphNetWithGIN. Our approach is trained and evaluated on the MPI-INF-3DHP dataset to ensure efficient and accurate 3D pose modeling. The method enables real-time processing while maintaining high precision, making it suitable for various applications requiring robust 3D pose estimation.