CVPR 2015

[ieee-7298683] Cascaded Hand Pose Regression [PDF] [notes] [dataset MSRA]

Xiao Sun, Yichen Wei, Shuang Liang, Xiaoou Tang, Jian Sun

read 29/05/2017

Synthesis

pose is parametrized by 21 kinematic joints

palm : 6 degrees of freedom in global viewpoint (3rotations, 3 translations)

global viewpoint is initialized as following :

hand models (distances between joints) are supposed known

Pipeline

Hand position initialization

• translation so that specific joint is at center of 3D point cloud of the hand

• rotation as PCA directions Y : larges variation (middle finger), Z : smallest variation (palm normal)

• joint initialized at their canonical positions

Iterative update of hand pose

• Each pose is refined using a regressor that takes as input the previous estimation and the depth image

• all training samples are transformed to the canonical coordinate frame using a full 3D rigid transformation, this defines a palm normalization transformation

• a finger normalization transformation is also defined for each finger

• first estimate palm and base of fingers, then estimate the fingers, each taking several iterations

Dataset

76.500 depth frames with 3D annotation (MSRA depth dataset)