3D human pose estimation with simple self-supervised learning

Abstract

Recent studies have shown remarkable advances in 3D human pose estimation from monocular images, with the help of large-scale in-door 3D datasets and ophisticated network architectures. However, the generalizability to different environments remains an elusive goal. In this work, we present a solution for single-view 3D human skeleton estimation based on deep learning method. Our network contains two separate model to fully regress and enhance the resulting poses. We utilize a newly proposed model whose name is Squeeze and Excitation Network (SE-net) as to construct our pose estimation network in order to estimate the corresponding pose from a color image; then a model consisting of several blocks of fully-connected networks and a novel semantic graph convolutional networks featuring self-supervision to reconstruct 3D human pose. We demonstrate the effectiveness of our approach on standard datasets for benchmark where we achieved comparable results to some recent state-of-the-art methods.