Structured Aleatoric Uncertainty in Human Pose Estimation

Human pose estimation from monocular images exhibits an inherent uncertainty through self-occlusions and interperson occlusions, aside from typical sources of uncertainty. Recently, there has been an increased focus in modelling uncertainty in supervised machine learning tasks. In line with this trend, we propose a novel formulation to capture aleatoric uncertainty in human pose using a multivariate Gaussian distribution over all the joints of human body and show that this improves generalization in 2D human pose estimation by implicitly suppressing the gradients from uncertain joints. Further, we develop a novel method to triangulate 3D human pose from predicted 2D poses, under the predicted uncertainty, that out-performs the baselines by over 10.8% and provide a multi-view inference benchmark for 3D human pose estimation on Human 3.6M dataset.