WGNet: Wider Graph Convolution Networks for 3D Point Cloud Classification with Local Dilated Connecting and Context-Aware
Yiping Chen,Zhipeng Luo,Wen Li,Haojia Lin,Abdul Nurunnabi,Yaojin Lin,Cheng Wang,Xiao-Ping Zhang,Jonathan Li
DOI: https://doi.org/10.1016/j.jag.2022.102786
IF: 7.5
2022-01-01
International Journal of Applied Earth Observation and Geoinformation
Abstract:Graph convolution networks (GCNs) have been proven powerful in describing unstructured data. Currently, most of existing GCNs aim on more accuracy by constructing deeper models. However, these methods show limited benefits, and they often suffer from the common drawbacks brought by deep networks, such as large model size, high memory consumption and slow training speed. In this paper, different from these methods, we widen GCNs to improve the descriptiveness by expanding the width of input to avoid the above drawbacks. Specifically, we present a wider GCNs based model, WGNet, for 3D point cloud classification. A local dilated connecting (LDC) module is designed to obtain the adjacency matrix, while a context information aware (CIA) module is proposed to generate initial node representation. These two modules provide a way to transform 3D point cloud into graph structure with larger receptive field and rich initial node features. These two properties widen the channels of input and provide more rich information to describe the samples precisely. Besides, we provide analysis to formulate the above idea as the sample precision description. Then, we adopt ChebyNet as our basic network, and present a skip-connection-based GCNs to improve efficiency of feature reuse. WGNet was evaluated on two datasets. One was acquired by a mobile laser scanning system under the real road environments, while the other was the well-known public artificial dataset, ModelNet40. Experimental results show that WGNet achieves better performance than the state-of-the-art in terms of descriptiveness, efficiency and robustness. Ablation studies also demonstrate the effectiveness of our designed LDC and CIA modules.