Abstract: Point clouds are widely used in various 3D application scenes. However, there are some issues in the real-world scenarios, such as labor-consuming and time-costing on data annotation. Few-shot point clouds classification neural networks are popular and can satisfy practical application requirements. To improve the performance of deep learning models in small sample point cloud classification, we propose a point clouds classification method based on MAML and Dirichlet process. To handle the imbalance of data sets, this paper proposes a similarity dependent Dirichlet process called Chinese restaurant model to cluster the dataset. This cluster process can complete clustering without manually specifying the number of clusters and result in higher classification accuracy with the optimal clustering. Subsequently, we apply MAML framework to train PointNet++ network based on samples processed by our cluster model. With the help of our data clustering, the MAML algorithm can adapt to new tasks quickly through a small number of data samples. The proposed method alleviates the issue of data hungry on deep learning for point clouds and improve the generalization of neural networks, as well as successfully extending the MAML algorithm from two-dimensional image classification to three-dimensional point clouds classification. Compared with PointNet++ model, the training time is reduced by half and the overall accuracy is improved by 6.67%. Experimental results demonstrate the effectiveness of our proposed small sample point cloud classification algorithm on public dataset.