Lie algebra-based multivalidation model for 3D human pose estimation

We investigate 3D stereo view human posture estimation. We investigate a straightforward architecture that uses intermediate 2D pose predictions to reason, in contrast to many methods that attempt to directly predict 3D stance from picture data. Our strategy is founded on two important findings. (1) 2D posture estimation has been transformed by deep neural networks, which can now provide precise 2D predictions even for poses with self-occlusions. (2) 2D stereo vision is fruitful enough to reconstruct 3D vision, making it tempting to “lift” predicted 2D poses to 3D through stereo triangulation. In this paper, we propose to model skeleton parts as the trajectory in the Lie group SE(n)×...×SE(n). The 3D pose estimation is performed based on the resulting trajectories mapped to a vector space, the Lie algebra, where the 3D pose of the human body is obtained by pose construction in Lie algebra space from a series of raw images from a different view, thus making sense how it reaches good performance. Our method is compared with other stereo-vision-based methods and outperforms them. Especially, our method shows good performance compared with the Microsoft Kinect sensor.