Fei-Fei Li-Lecture 16:3D Vision 【斯坦福大学李飞飞CV课程第16讲：3D Vision】

目录

P1 2D Detection and Segmentation​编辑

P2 Video = 2D + time series

P3 Focus on Two Problems

P4 Many more topics in 3D Vision

P5-10 Multi-View CNN 

P11 Experiments – Classification & Retrieval

P12 3D Shape Representations

P13--17 3D Shape Representations: Depth Map

 P18--26   3D Shape Representations: Surface Normals 曲面法线

P27--34  3D Shape Representations: Point Cloud

 P35--66    3D Shape Representations: Triangle Mesh

---

P1 2D Detection and Segmentation

Classification分P类：没有空间信息，只是对一张图片进行分类

Semantic Segmentation语义分割： 没有物体，只有像素点，对像素点进行分类

Object Detection目标检测：直接识别出物体并进行分类

Instance Segmentation：实例分割=目标检测+语义分割 （第一次听说这个）

语义分割只需要分出不同类就行，同类的不同个体不需要分，但是Instance Segmentation在语义分割的基础上又把不同的类进行了分割：目标检测后，需要对检测的部分做进一步的语义分割

P2 Video = 2D + time series

视频就是2D的图像加上了时间序列

P3 Focus on Two Problems

今天需要解决的两个问题

①由一张输入图像得到一个3D模型

②识别3D模型进行类别判定

P4 Many more topics in 3D Vision

3D Representations 三维表示法

Computing Correspondences 计算对应关系

Multi-view stereo 多视角立体

Structure from Motion 运动结构

Simultaneous Localization and Mapping (SLAM) 同步定位和绘图

View Synthesis 视图合成

Differentiable Graphics 可变图形

3D Sensors 三维传感器

P5-10 Multi-View CNN 

CNN1：提取图像特征的卷积神经网络

CNN2：生成描述形状符的卷积神经网络

P11 Experiments – Classification & Retrieval

Q：MVCNN? SPH? LFD? 3D ShapeNets? FV? 

P12 3D Shape Representations

Q： Voxel Grid?  Pointcloud?  Mesh?  Surface? 

A：下面详细讲啦

P13--17 3D Shape Representations: Depth Map

RGB image + Depth image = RGB-D Image (2.5D)

Q：H是Height ?  W是Width?

 可以使用全卷积神经网络进行深度图预测，得到两个估计的深度图，然后还可以得到每像素Loss 

Per-Pixel Loss (L2 Distance) 

Q：L2 距离是什么？

Problem: Scale / Depth Ambiguity

Q：具体是什么意思以及怎么解决？

A： 意思大概是单目图像中信息有限

Predicting Depth Maps

 Scale invariant 尺度不变性

 P18--26   3D Shape Representations: Surface Normals 曲面法线

对于每个像素，表面法线给出一个向量，表示该像素的世界上的对象的法向向量

假设 RGB Image为 3 x H x W，那么法线图 Normals: 3 x H x W

Predicting Normals

 3D Shape Representations: Voxels

• Represent a shape with a V x V x V grid of occupancies     网格表示形状

• Just like segmentation masks in Mask R-CNN, but in 3D!   分割掩码

• (+) Conceptually simple: just a 3D grid!     只是一个3D网格

• (-) Need high spatial resolution to capture fine structures 需要高空间分辨率捕捉精细结构

• (-) Scaling to high resolutions is nontrivial ！   缩放到高分辨率并不容易

Processing Voxel Inputs: 3D Convolution

 Generating Voxel Shapes: 3D Convolution

 Voxel Problems: Memory Usage

        Storing 1024（3次方） voxel grid takes 4GB of memory

Scaling Voxels: Oct-Trees 八叉树

Q： 没太看懂这个Oct-Trees

P27--34  3D Shape Representations: Point Cloud

• Represent shape as a set of P points in 3D space

• (+) Can represent fine structures without huge numbers of points

• ( ) Requires new architecture, losses, etc

• (-) Doesn’t explicitly represent the surface of the shape: extracting a mesh for rendering or other applications requires post-processing  

提取网格为渲染或其他应用提取网格需要进行后处理

Proessing Pointcloud Inputs: PointNet

 MLP ？  

Max-Pool？

Generating Pointcloud Outputs

 Predicting Point Clouds: Loss Function

 

 P35--66    3D Shape Representations: Triangle Mesh

 Predicting Meshes: Pixel2Mesh

 

 Idea #1: Iterative mesh refinement

Start from initial ellipsoid mesh Network predicts offsets for each vertex Repeat.

 Predicting Triangle Meshes: Graph Convolution

 

 Problem: How to incorporate image features?

Predicting Triangle Meshes: Vertex-Aligned Features

 

Predicting Meshes: Loss Function

The same shape can be represented with different meshes – how can we define a loss between predicted and ground-truth mesh?

Idea: Convert meshes to pointclouds, then compute loss

 3D Shape Prediction: Mesh R-CNN

Mesh R-CNN: Hybrid 3D shape representation

未完待续