3D Hand Pose Estimation from Single Depth Images with Label Distribution Learning

Reliable hand pose estimation enriches the way of human-computer interaction, such as sign language recognition and virtual reality. However, the task of estimating the hand pose faces two severe challenges. To be specific, it is difficult to learn spatial information from a 2D image and regress the location of a point in 3D space. And the highly non-linear correlation between the hand feature space and the joint location makes it hard to be modeled. To deal with the above problems, we propose a deep regression network, which learns the hand feature space from the point cloud and includes a specific label distribution learning network. Due to the point cloud contains more spatial information, it is beneficial for the neural network to extract the hand spatial geometric features. Utilizing the deep network to guide label learning actively reduces the negative effects of nonlinearity. According to the experimental results, our proposed network achieves the state-of-the-art performance on MSRA dataset.