Even though various studies have shown remarkable advancements in global human mesh reconstruction, most of those methods rely heavily on off-the-shelf SLAM algorithms or gyroscope measurements to transform mesh vertices from the camera coordinate to ...
Even though various studies have shown remarkable advancements in global human mesh reconstruction, most of those methods rely heavily on off-the-shelf SLAM algorithms or gyroscope measurements to transform mesh vertices from the camera coordinate to the world coordinate. To cope with this limitation, several approaches have been proposed to recover global human motions from local human poses, yet they still struggle to generate natural motions in complex scenarios. In this paper, we propose a simple yet effective method that predicts the global human trajectory to reconstruct temporally consistent 3D human meshes in the world coordinate based on a set of pre-defined directional vectors, called anchors. The key idea of the proposed method is to estimate the global trajectory by accumulating frame-wise translations, where each translation consists of two components, i.e., direction and magnitude. Specifically, each translation is estimated by determining the optimal direction selected from a set of anchors and then combining it with the magnitude value, which is learned through the proposed network. Such anchor-guided trajectory helps reduce noisy predictions by constraining the direction of the body translation to a set of plausible candidates. In addition, we propose to estimate relative 3D rotations with respect to a template orientation. This residual learning scheme is highly effective in capturing subtle changes in body orientations while reducing unexpected body rotations. Experiments on benchmark datasets show that the proposed method successfully improves the performance of 3D human mesh reconstruction in the world coordinate. Keyword : Anchor-guided trajectory, global human trajectory, video-based human mesh reconstruction.