Hand pose estimation plays an important role in human–computer interaction and augmented reality. Regressing the joints coordinates is a difficult task due to the flexibility of the joint, self-occlusion and so on. In this paper, we propose a novel and simple hierarchical neural network for hand pose estimation. The hand joint coordinates are divided into six parts and each part is regressed in sequence with this hierarchical architecture. This can divide the complex task of regressing all hand joints coordinates into several sub-tasks which can make the estimation more accurate. When regress the joint coordinates of one part, the features of other parts may bring negative influence to this part due to the similarity among the fingers, so we use an interference cancellation operation in our hierarchical architecture. At the time the joint coordinates of one part are regressed, the corresponding features will be removed from the hand global feature to eliminate the interference of this part. The obtained features will be used as input for regressing the joints coordinates of the next part. The ablation study verifies the effectiveness of our hierarchical architecture. The experimental results demonstrate that our method can achieve state-of-the-art or comparable results relative to existing methods on four public hand pose datasets.

手势估计在人机交互和增强现实中起着重要作用。由于关节的灵活性，自闭塞等原因，使关节坐标回归是一项艰巨的任务。在本文中，我们提出了一种新颖且简单的递阶姿势估计神经网络。手关节坐标分为六个部分，每个部分都按此层次结构顺序回归。这可以将使所有手部关节坐标回归的复杂任务分为几个子任务，这可以使估计更加准确。当回归某一部分的关节坐标时，由于手指之间的相似性，其他部分的特征可能对该部分带来负面影响，因此我们在分层体系结构中使用干扰消除操作。当一个零件的关节坐标回归时，相应的特征将从手全局特征中删除，以消除该零件的干扰。获得的特征将用作回归输入下一个零件的关节坐标的输入。消融研究验证了我们分层体系结构的有效性。实验结果表明，相对于四个公共手势数据集上的现有方法，我们的方法可以达到最新或相当的结果。