LGEFE: Effective Local-Global-External Feature Extraction for 3D Point Cloud Classification

Abstract

3D point cloud classification has a wide range of applications in computer vision, autonomous driving and robotics, and is of great importance. Unlike 2D image classification, the disorder and irregularity of 3D point clouds make it still more challenging to accomplish this task. In recent work, 3D point cloud classification still has some major shortcomings, including not fully utilizing the unique local spatial structure of 3D point clouds, global feature representation still has room for improvement, and not fully utilizing the relationship between different point cloud samples. To this end, this paper proposes an effective Local-Global-External Feature Extraction (LGEFE) neural network for 3D Point Cloud Classification. The feature extraction part of LGEFE consists of three blocks of Local (LFE), Global (GFE), and External (EFE) feature extraction, which use graph convolution, self-attention mechanism, and external attention mechanism to reasonably extract local, global, and external features of a single 3D point cloud sample, respectively. Extensive experimental results on the standard ModelNet40 dataset show that our proposed method exceeds the compared baseline, and ablation studies likewise validate the effectiveness of each of the proposed core components.