[Read the paper] DGCNN: Dynamic Graph CNN for Learning on Point Clouds

　　Complete set of network diagram into the pit, feeling a little difficult, a little slowly learn it, the proposed method is " Rethinking a using the Table Recognition Neural Networks Graph main method" in the modeling aspects of the relationship.

## Overview

　　This article is a classic PointNet improvements, the main objective is to design a point cloud can be used directly as an input CNN architecture can be applied to the classification, segmentation and other tasks. The main innovation is to propose a new micro network module can EdgeConv (convolution side) to extract local neighborhood information.

　　The overall network structure shown below, are worth noting:

• Similar to the structure of the network structure and PointNet overall, the most important difference is that instead of using EdgeConv MLP;
• EdgeConv for each module, i.e., we consider the overall characteristics, have to consider local features , aggregate functions  ;
  
• K value KNN is a block diagram EdgeConv hyperparametric classification network K = 20, and the dividing network K = 30;
• In the split network, each of the global descripter local descripter and after a connection point for each output prediction score;
• mlp after each feature is to calculate the edge (edge ​​features), dynamic convolution FIG.

  

## Edge Convolution

 

• The figure is a schematic diagram of EdgeConv. Suppose a dimensional point cloud F , wherein F represents the dimension of each point, the simplest possible be x, y, z three-dimensional, additional points may also be introduced into each color, normal information.
• 给定一个有向图 用来表示点云的局部结构，其中顶点为，边为 ，边特征函数，其中 h 是 
• 上图1就描述了一个点X_i和其邻近点X_j 的边特征e_ij求解过程，h使用全连接，用tf.layers.dense实现。(注：Dense and fully connected are two names for the same thing.)
• 特征聚合函数描述的是结点参数更新的过程，定义为□，其定义是：