Reproduce “TransPose: Real-time 3D Human Translation and Pose Estimation with Six Inertial Sensors” on OpenResearch

About this paper

Xinyu Yi, Yuxiao Zhou, Feng Xu

TransPose, developed by researchers at Tsinghua University, presents a real-time system for 3D human pose and global translation estimation using only six inertial sensors. It achieves over 90 frames per second, outperforming prior sparse IMU methods in accuracy and motion smoothness by up to an order of magnitude, while also providing global spatial tracking.

Key insights

• Decomposing the pose estimation problem into intermediate joint position predictions simplifies learning from sparse IMU data, leading to higher accuracy and smoother motions by better utilizing acceleration information.
• A hybrid approach fusing a supporting-foot assumption with a learned velocity predictor, coupled with an accumulative loss function, effectively addresses the challenges of real-time global translation from sparse IMUs.
• Training on large-scale synthetic IMU data generated from motion capture datasets, followed by fine-tuning on real IMU data, is crucial for developing robust deep learning models for under-constrained inertial motion capture problems.

Summary by alphaXiv.