Keypoint-Restored Inverse Kinematic Solver

Model-based 3D human body reconstruction provides a remarkable perspective for generating realistic full-body mesh in mesh-based 3D pose estimation. However, different strategies for capturing the nonlinear attributes of human model have varying effects on the efficiency and accuracy of recovering 3D human body. In this paper, our objective is to develop a simple and credible 3D human body parameter solver for skinned multi-person linear model (SMPL). To achieve this, we propose a keypoint-restored inverse kinematic solver (KRIKS), composed of inverse kinematics (IK) and iterative optimization, which transforms 3D keypoints to actual human shape and pose with mesh form. Additionally, considering the limitation of IK calculation and the relationship across pose frames, we extend our approach from single-frame analysis to video sequence in seq2frame manner. This integration leverages the temporal coherence of continuous actions, further enhancing the reconstruction performance. We demonstrate quantitative KRIKS baselines on the 3DPW and AMASS datasets, achieving impressive results with 0mm MPBE, 0.18mm PMPJPE, and 11.28mm MPVE on the 3DPW dataset.