Affinity-Point Graph Convolutional Network for 3D Point Cloud Analysis

Efficient learning of 3D shape representation from point cloud is one the biggest requirements in computer vision. In recent years, convolutional neural networks have achieved great success 2D image learning. However, unlike images that a Euclidean structure, clouds are irregular since neighbors each node inconsistent. Many studies tried to develop various graph overcome this problem and achieve results. Nevertheless, these simply took centroid its corresponding as thus ignoring structural information. paper, an Affinity-Point Graph Convolutional Network (AP-GCN) proposed learn structure for reference point. method, affinity between points first defined using feature feature. Then, with information built. After that, edge-conditioned convolution performed vertices edges obtain stronger neighborhood Finally, learned used recognition segmentation tasks. Comprehensive experiments demonstrate AP-GCN much more reasonable features significant improvements vision tasks such object classification segmentation.