3D human pose estimation with temporal and position enhancement

Abstract

The existing 3D human pose estimation method in exploring the relationship between root joint and other key points, to some degree reflects the local movement, but they ignore the trajectories of the whole of the human body, and to some extent, it reflects the global movement, even explore the method of the root node and other related points made significant progress, However, it is not robust to global motion, and it is not perfect to small local motion. For the above two kinds of situations, we proposed representation of temporal and position enhancement. Position enhancement is to maintain the consistency of the input and output distribution by position encoding the correlation of 2D coordinates of human pose. Temporal enhancement occurs by making connections between the current posture of the same person over time and other postures, as we focus more on the current posture and changes in motion in the forward and backward posture.