This study aims to propose the first unsupervised deep learning based framework to detect primitive shapes in unorganized point clouds. The benefits of the proposed model are as follows: first, modeling primitive shapes mathematically and making the annotation and generation of training data automatically can assist in training a network in an unsupervised manner without using any pre-annotated datasets. The proposed approach is able to expand to other 3D point clouds related tasks when the datasets are not big enough. Experiments and results demonstrate that the proposed unsupervised methods can achieve an accuracy on par with supervised approach for primitive shape detection in 3d point clouds.