A Relation Network Based Approach for Few-Shot Point Cloud Classification

As a commonly used format of 3D data, point clouds preserve the original geometric information in 3D space without any discretization. In recent years, many deep learning methods have been proposed for recognizing and classifying 3D point cloud data. These methods often require a large number of labeled point clouds for training. However, it is obviously difficult to obtain enough labeled samples for all classes of point clouds in practice. To address this issue, this paper proposes a relation network based on point cloud classification method which can recognize the objects that the point cloud data represents with only few labeled samples. In order to better obtain the local neighborhood information, we use EdgeConv operator to extract the features of each point of the point clouds. And the class of a point cloud will be predicted by measuring the similarity between its feature and the prototypes of a few marked point clouds. Based on the dataset of ModelNet40, the experiments have shown that the proposed method can achieve 92.48% in accuracy and shows better performance compared with related works.