3D Pose Prediction: Unsupervised Adversarial Network With Human Geometric Constraint

2D human pose extraction has been completed by multiple frameworks, such as [1], [2], [3], [5], [6], [7], [8], [9], [1]0,[1][1] and other detectors. In this paper, after comparing various 2D detectors, 1.7 version of openpose, the most cutting-edge of CMU, is adopted as the 2D pose detector of this paper for accuracy. As for the transformation from 2D joint pose to 3D pose, Pavllo et al. [4] was referred to, while some modifications were made to the errors still existing in the actual output results. We use the classic GAN[23] model, and the network adopts unsupervised training mode. On this basis, We have added a constraint network to the network layer design to make the 3D human pose predicted by the original confrontation generation network obtain a higher accuracy, and in addition to the generation network we use strong geometric constraints on the key points of the human body which cannot be trained on the network(Such as elbows and knees), the entire human bones are contracted in millimeters, so that the predicted 3D posture will not be too distorted. Experiments have proved that such a method can obtain a state-of-the-art low-error result. The unsupervised network structure makes the framework not dependent on the huge annotated 3D data set. We only use the 3D annotation dataset in the final verification part, the 3D pose data set is provided with real human body pictures from the official website of Human3.6m [13]. Finally, the effective 3D prediction results were projected into the skeleton action Data of Vocaloid Motion Data(VMD)[37]and used to obtain The 3D animation corresponding to 2D poses in the Skeleton editing engine of MikuMikuDance[38].