Linked Dynamic Graph CNN: Learning through Point Cloud by Linking Hierarchical Features
Kuangen Zhang, Ming Hao, Jing Wang, Xinxing Chen, Yuquan Leng, Clarence W. de Silva, Chenglong Fu
Abstract
Learning through a point cloud is attractive be-cause a point cloud contains geometric data and can help robots understand environments in a robust manner. However, a point cloud is sparse, unstructured and unordered, and its accurate recognition by a traditional convolutional neural network (CNN) or a recurrent neural network (RNN) is difficult. Hence, the present paper proposes a linked dynamic graph CNN (LDGCNN) to directly classify and segment a point cloud. The present work removes the transformation network, links hierarchical features from dynamic graphs, freezes the feature extractor, and retrains the classifier in order to increase the performance of LDGCNN. Also, the paper describes LDGCNN using theoretical analysis and visualization. The LDGCNN is able to classify a point cloud consisting of 1024 three dimensional points without normal vectors in the ModelNet40 dataset, achieving 92.9% accuracy, which is better than the state-of-art. The LDGCNN also achieves state-of-the-art segmentation performance in the ShapeNet dataset.