Uncertainty-Aware Adaptation for Self-Supervised 3D Human Pose Estimation

The advances in monocular 3D human pose estimation are dominated bysupervised techniques that require large-scale 2D/3D pose annotations. Suchmethods often behave erratically in the absence of any provision to discardunfamiliar out-of-distribution data. To this end, we cast the 3D human poselearning as an unsupervised domain adaptation problem. We introduce MRP-Netthat constitutes a common deep network backbone with two output headssubscribing to two diverse configurations; a) model-free joint localization andb) model-based parametric regression. Such a design allows us to derivesuitable measures to quantify prediction uncertainty at both pose and jointlevel granularity. While supervising only on labeled synthetic samples, theadaptation process aims to minimize the uncertainty for the unlabeled targetimages while maximizing the same for an extreme out-of-distribution dataset(backgrounds). Alongside synthetic-to-real 3D pose adaptation, thejoint-uncertainties allow expanding the adaptation to work on in-the-wildimages even in the presence of occlusion and truncation scenarios. We present acomprehensive evaluation of the proposed approach and demonstratestate-of-the-art performance on benchmark datasets.