Adapted human pose: monocular 3D human pose estimation with zero real 3D pose data

The ultimate goal for an inference model is to be robust and functional in real life applications. However, training vs. test data domain gaps often negatively affect performance. This issue especially critical the monocular 3D human pose estimation problem, which collected a controlled lab setting. In this paper, we focus on alleviating negative effect of shift both appearance space by presenting our adapted (AHuP) approach. AHuP built upon two key components: (1) semantically aware adaptation (SAA) cross-domain feature adaptation, (2) skeletal (SPA) takes only limited information from target domain. By using zero data, one synthetic models shows comparable performance with SOTA trained large scale datasets. proposed SPA can also employed independently as light-weighted head improve existing novel context. A new scan-based dataset called ScanAva+ going publicly released work.