Unsupervised Reconstruction of 3D Human Pose Interactions From 2D Poses Alone

• DOI: 10.48550/arxiv.2309.14865
• Publication Date: 2023-01-01
• Authors: P. Hardy; Hansung Kim
• Keywords: 3D Reconstruction; Benchmark (surveying); Monocular

Current unsupervised 2D-3D human pose estimation (HPE) methods do not work in multi-person scenarios due to perspective ambiguity monocular images. Therefore, we present one of the first studies investigating feasibility HPE from just 2D poses alone, focusing on reconstructing interactions. To address issue ambiguity, expand upon prior by predicting cameras' elevation angle relative subjects' pelvis. This allows us rotate predicted be level with ground plane, while obtaining an estimate for vertical offset 3D between individuals. Our method involves independently lifting each subject's 3D, before combining them a shared coordinate system. The are then rotated and being scaled. itself enables retrieve accurate reconstruction their poses. We our results CHI3D dataset, introducing its use three new quantitative metrics, establishing benchmark future research.