What

DIP is a Deep-Learning model predicting full body pose from sparse (only 6) Inertial-Measurement-Units (IMUs) in real-time. DIP is firstly trained on a large scale synthetic dataset generated from archival MoCap data, then fine-tuned on DIP-IMU, the real IMUs data paired with SMPL pose reference we built ourself. DIP can be used in various application scenarios, especially AR and VR.


Authors

Yinghao Huang, Max Planck Insitute for Intelligent Systems, Tübingen

Manuel Kaufmann, Advanced Interactive Technologies Lab, ETH Zürich

Emre Aksan, Advanced Interactive Technologies Lab, ETH Zürich

Michael J. Black, Max Planck Insitute for Intelligent Systems, Tübingen

Otmar Hilliges, Advanced Interactive Technologies Lab, ETH Zürich

Gerard Pons-Moll, Max Planck Institute for Informatics, Saarbrücke