PointNet学习笔记

PointNet学习笔记

PointNet以及PointNet++源码

PointNet++ 点云处理原理

三维点云技术概述

3D representation

• Point Cloud

• Mesh

• Volumetric

• Projected View（RGB-D）

2.5D RGB-D Depth Map

3D Point Clouds

点云应用

• Robot Rerception

• Augmented Reality（AR）

• VR

• Shape Design

• FaceID

点云处理任务

• Oject classification 物体分类

• Parts segmentation 部件分割

• Object detection 目标检测

• Semantic segmentation (场景语义分割)

数据集

![在这里插入图片描述](https://img-blog.csdnimg.cn/20210320

Challenges

• Irregular.近密远疏
• Unstructured
• Unordered
  • Invariance to permutations 置换不变性

Structured grid based learning

• Voxel based

• multiview based

• Point-Based Methods

  • PointNet 开山之作 MLPs --> Max pooling
  • PointNet++ Sampling --> Grouping

• Convolution-based Methods

  • 3D neighboring points
  • 3D continuous convolution
  • 3D discrete convolution

• Graph-based Methods

  • Input points --> Graph Construnction --> Feature Learning & Pooling --> Output Points

A taxonomy of deep learning methods for 3D point clouds

Chronological overview

PointNet 点云处理原理

Previous Works

Point clud is converted to other representations before it’s fed to ad deep neural network

Conversion	Deep Net	Algorithm	Shortcoming
Voxelization	3D CNN	VoxNet	栅格量化以后分辨率会降低，存在信息丢失
Projection/Rendering	2D CNN	Multi-view CNN	Rendering个数增加计算量也随之增加
Feature etraction	Fully Connected

PointNet

feature learning directly

• End-to-end learning for scattered, unordered point data
• Unified framework for various tasks

Challenges

• Unorder point set as input

  Model needs to be invariant to N! permutations

• Invariance under geometric transformations

  Point cloud rotations should not alter classification results

PointNet(vanilla)

Universal Set Function Approximator

PointNet的网络结构能够拟合任意的连续集合函数

作者证明Max pooling的引入不会降低拟合其他函数的能力（通过将特征映射到高维）

Input Alignment by Transformer Network

Idea: Data dependent transformation for automatic alignment

Embedding Space Alignment

PointNet Classification Network

Extension to PointNet Segmentation Network

Model Size and Speed – Light-weight & Fast

PointNet++ 点云处理原理

PointNet没有local context

Basic idea

Recursively apply pointnet at local regions

• Hierarchical feature learning

• Translation invariant

• Permutation invariant

Hierarchical feature learning

Set Abstraction：sampling + grouping + pointnet

PointNet++ PointNet++ 总体思路 : 首先通过将点集划分为重叠的局部区域。 与CNN 相似，提取 局部特征以捕获来自小邻域的精细几何结构。 这些局部特征将进一步分组为较大的单 元，并进行处理以生成更高级别的特征。 重复此过程，直到获得整个点集的特征为止。

• Sampling

  • Uniform sampling
  • Farthest sampling

• Grouping

  • K nearest neighbors
  • Ball query（within range)

• Apply PointNet to each group
  

对于非均匀点云的处理方法

• Multi-scale grouping（MSG）
• Multi-resolution grouping(MRG)

Network Architectures

• SSG:single scale grouping
• FP（feature propagation）

Results: Non-Euclidean Space

卷积感受野

在卷积神经网络中，感受野（Receptive Field）的定义是卷积神经网络每一层输出的特征图（feature map）上的像素点在输入图片上映射的区域大小。再通俗点的解释是，特征图上的一个点对应输入图上的区域。