Fitting a deformable 3D human body model to depth images using convolutional neural networks

Abstract

This work combines two existing approaches for 3D human shape completion. A generative statistical model of human shape (SCAPE) and a correspondence algorithm based on convolutional neural networks. The correspondences are used to control a nonrigid iterative closest points (NICP) algorithm which is regularized by a SCAPE model. We expect that this approach will mitigate the initialization problem of NICP and detect correspondence mismatches of a trained feature extractor.