Hand Pose Ensemble Learning Based on Grouping Features of Hand Point Sets

In this paper, we mainly consider using 3D point sets as input to deal with the task of 3D hand pose estimation. We make some improvements to PointNet++ structure, including proposing adaptive pooling which introduces the self-attention mechanism to make the network could select features itself, and putting forward an ensemble strategy to fully utilize hand features. These improvements can enhance the expressive ability of features and make full use of the information contained in features. In addition, we propose a data augmentation method for point net, which directly transforms the original point cloud data without the aid of simulation models. Experiments results on three hand pose datasets demonstrate that our method can achieve comparable performance with state-of-the-arts.