Dual-Attention Network with Aggregate Local Descriptors Vector for 3D Point Cloud Classification

Contextual fine-grained local features play an essential position in 3D point cloud classification, and have not been utilized effectively in existing deep-learning-based models. Aiming to address this problem, a 3D point cloud classification network based on a dual attention mechanism with vector of locally aggregated descriptors(VLAD) is proposed. First, the neighbourhood fine-grained features and global information of the point cloud are mined by using a self-attention mechanism, and then the local geometric representation is learned by embedding a graph attention mechanism in a multilayer perceptron layer. To make the most of the features, a multiheaded mechanism is applied to aggregate different features from separate headers, and an effective key point descriptor is introduced to help identify the global geometry. Finally, the high-level semantic features of point clouds are obtained by the VLAD layers. Through extensive experiments, the model achieves 92.45% accuracy on the ModelNet40 dataset.