Point Clouds Learning Using Directed Connected Graph

Abstract

With the development of various 3D sensors, it has become easier for humans to obtain information in the 3D world, more and more people turn their attention to the problem of point clouds understanding. At present, most of methods focus on directly extracting features from point clouds, where feature extraction is performed by Multi-Layer Perception (MLP) and fusion is by local pooling. However, they do not consider the spatial relationship within the local point sets. We propose a directed connected graph network (DCGN), which can effectively capture the spatial relationship of local point sets by constructing the directed connected graph (DCG). Specifically, in the feature learning stage, the connection directions from neighbor points to the center point are constructed for each local point set to learn the feature transferring weights from neighbor points to center point. In order to further model the spatial distribution of local point sets, we use a distance-weighted method to perform local feature fusion. Extensive experimental results demonstrate that our method can achieve competitive performance on some standard data sets.