3D Hand Shape and Pose Estimation

Abstract

We study the problem of 3D hand shape and pose estimation from monocular RGB images. Recent studies have shown that single-view 3D hand pose estimation is challenging due to depth ambiguity, environmental conditions, object-occlusion, and self-occlusion. Further, acquiring 3D annotations for datasets requires significant efforts. In this research, towards solving these challenges, we propose some contributions. First, we address the problem of estimating the 3D hand shape and pose in a video dataset given only sparsely annotated frames. We propose label propagation to propagate 3D annotations from labelled frames to nearby unlabelled frames. Next, we address the problem of probabilistic 3D hand shape and pose estimation. Most existing works only estimate a unique solution for a 2D observation and ignore depth and occlusion ambiguities. In contrast, we learn a probability distribution over the hand parameters to generate multiple hypotheses for a given 2D observation. Show more