IENET: Implicit and Explicit Graph Network for Point Cloud Analysis

Since 3D point clouds have complex geometric information and it is difficult to extract truly meaningful local features, point cloud analysis has been a hot research topic in recent years. In this paper, we propose an implicit and explicit graph network (IENet). Specifically, IENet differs from other methods in that it can not only extract implicit features of point neighborhoods using adaptive convolution, but also obtain explicit features of neighborhoods based on the relative position information between points. Thus IENet combines the advantages of adaptive kernel and attention to better extract the local information of point clouds. We have carried out experiments of the proposed model on the point cloud classification dataset ModelNet and the point cloud part segmen-tation dataset ShapeNet Part, and the results both demonstrate the advancedness of the method.