Deep Neural Network for Point Sets Based on Local Feature Integration

Abstract

The research of object classification and part segmentation is a hot topic in computer vision, robotics, and virtual reality. With the emergence of depth cameras, point clouds have become easier to collect and increasingly important because of their simple and unified structures. Recently, a considerable number of studies have been carried out about deep learning on 3D point clouds. However, data captured directly by sensors from the real-world often encounters severe incomplete sampling problems. The classical network is able to learn deep point set features efficiently, but it is not robust enough when the method suffers from the lack of point clouds. In this work, a novel and general network was proposed, whose effect does not depend on a large amount of point cloud input data. The mutual learning of neighboring points and the fusion between high and low feature layers can better promote the integration of local features so that the network can be more robust. The specific experiments were conducted on the ScanNet and Modelnet40 datasets with 84.5% and 92.8% accuracy, respectively, which proved that our model is comparable or even better than most existing methods for classification and segmentation tasks, and has good local feature integration ability. Particularly, it can still maintain 87.4% accuracy when the number of input points is further reduced to 128. The model proposed has bridged the gap between classical networks and point cloud processing.