In recent papers, mixture of experts is used to overcome the ambiguities occurred in 3D human pose estimation from monocular images or videos. However, because of the high dimension of the image and pose space, a large amount of labeled samples are required during estimation, i.e. images with their corresponding poses, this demands considerable human effort. In this paper, we use a semi-supervised style that utilizes both labeled and unlabelled samples to deal with the task. Manifold regularization is introduced as prior information to pilot each expert. Experimental results in real image sequences illustrate that our framework truly works well.