Lifting 2D Human Pose to 3D with Domain Adapted 3D Body Concept

Lifting the 2D human pose to 3D is an important yet challenging task. Existing estimation suffers from (1) inherent ambiguity between and data, (2) lack of well-labeled 2D–3D pairs in wild. Human beings are able imagine a image or set body key-points with least ambiguity, which should be attributed prior knowledge that we have acquired our mind. Inspired by this, propose new framework leverages labeled poses learn concept reduce ambiguity. To consensus on pose, key insight treat as two different domains. By adapting domains, learned applied guides encoder generate informative “imagination” embedding lifting. Benefiting domain adaptation perspective, proposed unifies supervised semi-supervised principled framework. Extensive experiments demonstrate approach can achieve state-of-the-art performance standard benchmarks. More importantly, it validated explicitly effectively alleviates improves generalization, enables network leverage abundant unlabeled data.