This paper provides a novel method called "PointCNN-Hand" for 3D hand joints estimation based on PointCNN. To use the depth image effectively, we transfer the hand depth image into the 3D hand cloud point and implement end-to-end training by PointCNN-Hand for hand joint estimation. We then perform error analysis on MSRA, NYU, and ICVL datasets to compare with the state-of-the-art methods. The experiments show that the proposed method has desired results, and the model parameters are relatively smaller than those of other methods. To be specific, the parameters of the proposed PointCNN-Hand network are reduced to only 3 Mega Byte (MB) with Floating Point Operations (FLOPs) less than 232.05M.