Estimating 3D body mesh without SMPL annotations via alternating successive convex approximation

Abstract

This paper focused on extracting effective human shape and pose information from solely joint annotations. With few training datasets and without using SMPL annotations from MoSh, we proposed a method based on alternating successive convex approximation (ASCA) to estimate the 3D human shape and pose. Previous methods tended to utilize a large number of mixed 2D and 3D datasets for training. These methods extract shape and pose information from SMPL annotations containing ground-truth shape and pose annotations. It is challenging to learn useful information of human shape and pose from solely joint annotations independently because 2D and 3D joint positions can be expressed as a non-convex function of coupled SMPL shape and pose parameters. The proposed method decouples the function into joint-shape and joint-pose functions making the training focus on shape and pose separately during each procedure. A minimum number of datasets are used to train the proposed method (InstaVariety and MPI-INF-3DHP) compare to previous methods. We make a comparison with other methods that have been trained with SMPL annotations. The result shows that the proposed method is competitive with other algorithms trained with additional SMPL annotations. What is more, when ground-truth SMPL annotations exist, ASAC can equally extract useful shape and pose information from image sequences. The result trained with a 3DPW training set outperforms most current video-based algorithms.