SVMAC: Unsupervised 3D Human Pose Estimation from a Single Image with Single-view-multi-angle Consistency

Recovering 3D human pose from 2D joints is still a challenging problem,especially without any 3D annotation, video information, or multi-viewinformation. In this paper, we present an unsupervised GAN-based modelconsisting of multiple weight-sharing generators to estimate a 3D human posefrom a single image without 3D annotations. In our model, we introducesingle-view-multi-angle consistency (SVMAC) to significantly improve theestimation performance. With 2D joint locations as input, our model estimates a3D pose and a camera simultaneously. During training, the estimated 3D pose isrotated by random angles and the estimated camera projects the rotated 3D posesback to 2D. The 2D reprojections will be fed into weight-sharing generators toestimate the corresponding 3D poses and cameras, which are then mixed to imposeSVMAC constraints to self-supervise the training process. The experimentalresults show that our method outperforms the state-of-the-art unsupervisedmethods on Human 3.6M and MPI-INF-3DHP. Moreover, qualitative results on MPIIand LSP show that our method can generalize well to unknown data.