Single-Image 3D Human Pose and Shape Estimation Enhanced by Clothed 3D Human Reconstruction

3D human pose and shape estimation and clothed 3D human reconstruction are two hot topics in the community of computer vision. 3D human pose and shape estimation aims to estimate the 3D poses and body shapes of "naked" humans under clothes, while clothed 3D human reconstruction refers to reconstructing the surfaces of humans wearing clothes. These two topics are closely related, but researchers usually study them separately. In this paper, we enhance the accuracy of the 3D human pose and body shape estimation by the reconstructed clothed 3D human models. Our method consists of two main components: the 3D body mesh recovery module and the clothed 3D human reconstruction module. In the 3D body mesh recovery module, an intermediate 3D body mesh is first recovered from the input image by a graph convolutional network (GCN), and then the 3D body pose and shape parameters are estimated by a regressor. In the clothed human reconstruction module, two clothed human surface models are respectively reconstructed under the guidance of the recovered 3D body mesh and the ground-truth 3D body mesh. At the training phase, losses which are described by the residuals among the two reconstructed clothed human models and ground truth are passed back into the 3D body mesh recovery module and used for boosting the body mesh recovery module. The quantitative and qualitative experimental results on THuman2.0, and LSP show that our method outperforms the current state-of-the-art 3D human pose and shape estimation methods.