SparsePoser: Real-time Full-body Motion Reconstruction from Sparse Data

  The paper “SparsePoser: Real-Time Full-Body Motion Reconstruction from Sparse Data” introduces SparsePoser, a deep learning-based method for reconstructing high-quality full-body human poses in real time using only a small number of 6-degree-of-freedom (6-DoF) tracking sensors. Traditional motion capture systems and many IMU-based approaches require many sensors or struggle with positional ambiguity and drift, but SparsePoser combines a convolutional autoencoder that learns the manifold of human motion with a learned inverse kinematics (IK) component to refine hand and feet placement relative to the sparse trackers. Extensive evaluation on publicly available motion capture datasets and live demonstrations shows that SparsePoser produces continuous, natural motion that outperforms state-of-the-art techniques while operating with low latency, making it suitable for VR/AR, gaming, and animation applications where cost-effective and accurate motion tracking is essential. Read it here.