Three dimensional (3D) human model acquisition using single-viewpoint provides flexible setup with a trade-off of partial point clouds. By leveraging the learning capability of neural networks, a complete point clouds can be obtained by sampling a reconstructed 3D human model. A skeletal joints-based regressive 3D human reconstruction is proposed in this paper to infer skeletal joints and variance for reconstructing human shape. A skeletal joints encoder is proposed to learn the latent representation of input partial point clouds using Gaussian maximum likelihood to obtain localized skeletal joints and its variances. The skeletal joints are fed to a regressive human model to reconstruct a synthetic human model to obtain a complete point cloud. Lastly, a two-mode training strategy is proposed to enhance the learning of the proposed method by employing synthetic training in prior and non-synthetic fine-tuning. A real human motion dataset is used in the experiment as the performance evaluation to study the skeletal joints estimation and human shape learning.