Multi-View Human Pose Estimation with Geometric Projection Loss

3D Human Pose Estimation (HPE) has emerged as a significant area of focus, with triangulation being a pivotal technique for multi-view pose estimation, valued for its efficiency and effectiveness. Traditional approaches, including both supervised and semi-supervised triangulation methods, typically necessitate substantial volumes of 3D labeled data, the acquisition of which is challenging in practical scenarios. This paper introduces a novel unsupervised triangulation method for estimating 3D keypoints that leverages the inherent geometric properties of the triangulation process. Specifically, the method involves calculating the Euclidean distance between the triangulated points and their corresponding projection rays, coupled with a novel scoring mechanism for each view. By integrating consistency constraints and global contextual information, we refine our triangulation process to enhance accuracy. Extensive evaluations on the Human 3.6m dataset demonstrate that our method outperforms other baseline methods and significantly improves the accuracy of triangulation.