3D Human Pose Estimation in the Wild by Adversarial Learning 

Figure 1. Given a monocular image and its predicted 3D pose, the human can easily tell whether the prediction is anthropometrically plausible or not (as shown in b) based on the perception of imagepose correspondence and the possible human poses constrained by articulation. 

We simulate this human perception by proposing an adversarial learning framework, where the discriminator is learned to distinguish ground-truth poses (c) from the predicted poses generated by the pose estimator (a, b), which in turn is enforced to generate plausible poses even on unannotated in-the-wild data.