Read the paper "GAPNet: Graph Attention-based Point Neural Network for Exploiting Local Feature of Point Cloud

《GAPNet: Graph Attention-based Point Neural Network for Exploiting Local Feature of Point Cloud》

research problem	How to better extract the local feature point clouds
The main idea	And using the self-attention neighboring attention mechanisms, on the one hand its input points extracted by the geometric features self-attetion, on the other hand to extract the relationship between the input node and neighbors by neighboring atteion Using multi-head manner to obtain the original point cloud richer semantic feature
solution	self_attetion implementation: By mapping the first MLP each feature point to a higher-dimensional space, then the channel characteristic points of a weighted sum of each point to obtain further global feature neighboring achieve atteion of: DGCNN similar manner, in order to let each point of the original point cloud constructed as a central point corresponding kNN graph, and then further kNN graph based feature extraction, so as to obtain a local feature point for each local weights and graph attention： By self-attetion we can get right to the center point of each weight, we can get through neighboring attention right around the center point of each edge feature heavy, and finally we have to do is attetion weights based on the information side of the information and points of integration, Further features of the resulting final total Multi-head attention: Parallel repeat the process above graph attention in order to get richer geometrical information.