[点云目标检测]-A Hierarchical Graph Network for 3D Object Detection on Point Clouds

A Hierarchical Graph Network for 3D Object Detection on Point Clouds

CVPR 2020

摘要

原文	译文
3D object detection on point clouds finds many applications.	3D点云物体检测有越来越多的应用.
However, most known point cloud object detection methods did not adequately accommodate the characteristics (e.g., sparsity) of point clouds, and thus some key semantic information (e.g., shape information) is not well captured.	但是，目前已知的点云目标检测方法并没有充分适应点云的特性(例如，稀疏性)，而且一些关键的语义信息(例如，形状信息)没有很好的捕捉到.
In this paper, we propose a new graph convolution (GConv) based hierarchical graph network (HGNet) for 3D object detection, which processes raw point clouds directly to predict 3D bounding boxes.	本文提出一种新的适用于3D目标检测的图卷积网络结构，可以直接处理原始点云数据来预测3D框。
HGNet effectively captures the relationship of the points and utilizes the multi-level semantics for object detection. Specially, we propose a novel shape-attentive GConv (SA-GConv) to capture the local shape features, by modelling the relative geometric positions of points to describe object shapes.	特别的，提出了一种新的形状注意力图卷积算子(SA-GConv)，来捕捉局部形状特征，通过对点之间的相对几何位置进行建模来描述物体的形状。
An SA-GConv based U-shape network captures the multi-level features, which are mapped into an identical feature space by an improved voting module and then further utilized to generate proposals.	提出一个基于SA-GConv U形网络来获取多层特征，并且提出一个改进的voting模块来产生proposals.
Next, a new GConv based Proposal Reasoning Module reasons on the proposals considering the global scene semantics, and the bounding boxes are then predicted.	然后，提出了一个基于GConv的Proposal reasoning模块来对proposal进行reason，考虑全了局语义。
Consequently, our new framework outperforms state-of-the-art methods on two large-scale point cloud datasets, by ∼4% mean average precision (mAP) on SUN RGB-D and by ∼3% mAP on ScanNet-V2.	在SUN RGB-D 和ScanNet-V2上尽进行了实验，超过了SOTA.

Shape attentive Graph Convolutions

重点看一下这个shape-attentive模块怎么玩的，感觉创新点不是很大哎。
首先，作者给了下面这个公式和一个示意图

一个是g函数，一个是f函数。g就是描述shape的关键，输入是点i和点j的相对位置即$(x_i-x_j,y_i-y_j,z_i-z_j)$，将3维变成1维，作者说是用a simple one-by-one convolution with the Sigmoid activation function.然后f就是常见的MLP，输入是点i和点j的特征。然后将g的输出对应的乘以f的输出，得到中心点的knn，最后用一个maxpool得到点i的新特征。

虽然形式上没有attention中的softmax这样的归一化，但是g的输出就和attention一样，每个点的weights，然后对应的乘以特征。大概这就是shape attentive的意思吧。
在实验部分，作者比较了SA-GConv和GConv，SA能有2%的提升，说明还是很有用的。
并且还有可视化，说明g模块对edge, corners, sharp parts有高响应。