Local and Global Point Cloud Reconstruction for 3D Hand Pose Estimation

The 3D shape and pose of the human hand are critical for augmented and virtual reality applications. To accommodate this form of human-computer interaction, an entire discipline of computer vision is devoted to estimating 3D hand shape and pose. Achieving accurate estimates is extremely challenging due to the hand's high degrees of articulation and self-occlusion. Earlier approaches attempted to combine representations from various viewpoints [7, 8, 9, 10], or transform 2.5D depth maps to 3D representations such as voxels [23, 25], point clouds [21], or meshes [34]. Since 3D voxel models are computationally more expensive than mesh and point cloud models, the latter two are preferable for estimating 3D hand shape and pose. Current RGB-based methods [2, 19, 46] prefer to estimate hand shape by mapping visual features to the parameters of a parametric model e.g.