3D Human Pose Estimation Based on Volumetric Joint Coordinates

Estimating three-dimensional human pose from color images of single person is a fundamental problem in many applications.However,the problems of inaccuracy and ill-posed poses have been not well solved.A novel deep learning based approach for estimating 3D human poses from images is proposed.First voxel representation is adopted and the joint coordinates are presented to represent the poses.Second,space integral regression is used to compute the output results of the convolutional network.Finally,the output is sent into the fully connected network for joint training.Proposed algorithm has been tested under two standard test protocol of human3.6m dataset.Experimental results show that it obtains higher accuracy than most of previous methods and achieves well generalization ability in MPI-INF-3DHP dataset.