DPPCN: density and position-based point convolution network for point cloud segmentation

A point cloud can usually describe the outline and spatial location of an object. Due to the disorder and uneven density of the point cloud, it is a difficult task to fully obtain the local features and spatial context information of the point cloud. In this paper, we propose a point cloud segmentation network based on the encoding-decoding structure of point convolution, which extracts the local features of point clouds by density-position adaptive convolution, which integrates density information and positional relationships between points. To obtain the density information of center points, we design an auto-adjusted bandwidth and integrate it into adaptive kernel density estimation. In addition, to obtain the context of the point cloud to a greater extent, we design an encoding layer that carries the contextual information. In order to verify the effectiveness of our method, experiments were carried out on S3DIS and a self-built dataset. The experimental results verify the validity of our proposed method.