3D Human Mesh Recovery with Learned Gradient

Abstract 3D human body reconstruction is a research hotspot in computer vision. In this article, we propose a method for 3D human mesh recovery with learned gradient. We use gradient descent networks to predict model parameter updates. Previous studies have focused on 3D pose estimation, often ignoring the importance of body shape reconstruction. In order to improve the accuracy of human body model shape estimation, we introduce silhouette information that can represent human body shape, using silhouette and 2D joints as inputs to the neural network. In addition, throughout the entire training process, we only need the SMPL parameterized human pose dataset, without any image to 3D correspondence. The network learns effective pose and shape subspaces from this data, and performs optimization more effectively in that subspace. It has been proven that our method can achieve advanced performance on public datasets. Even on challenging datasets in the wild, competitive results can be obtained.