本文共 1332 字，大约阅读时间需要 4 分钟。

吴恩达 Andrew Ng

目标定位 Object Localization

• Classification with localization 分类并且确定目标位置（一个物体）
  
• Detection 定位（多个物体）
• target label y
  

特征点检测 Landmark Detection

• 设置特征点坐标值作为输出
• 所有标签在图片中保持一致 labels are consistent across different images

目标检测 Object Detection

• Sliding windows detection 滑动窗口检测
  • 固定步幅，滑动窗口，遍历图像的每个区域
  • 裁剪后的图像输入卷积网络，对每个位置进行分类
  • 卷积网络进行单个分类的计算成本很高
• granularity 粒度，stride 步幅

Convolutional Implementation of Sliding Windows

• Turing FC layer into convolutional layers
  

• Convolution implementation of sliding windows

  

Bounding Box Predictions

• YOLO algorithm (You Only Look Once), more accuracy bounding box
• assign an object to grid cell which contains the mid point of the object

交并比 Intersection Over Union

• 衡量定位精确度的一种方式
• 交集除以并集
• a measure of the overlap between two bounding boxes

非极大值抑制 Non-max Suppression

• output maximal probabilities classifications
• suppress closed and non-maximal ones
• 把其他与最大概率的边界框有高交并比的边界框抑制掉
• 

Anchor Boxes

• 一个格子中检测两个物体
• Each object in training image is assigned to grid cell that contains object’s midpoint and anchor box for the grid cell with highest IoU
• 两个物体的中心的位于同一个格子概率很小
• 人工选择 Anchor Box 的形状来配合需要检测的物体
• 通过 k-means algorithm 聚类，选择合适的形状

YOLO Algorithm

Training

Making predictions

Output the non-max supressed outputs

候选区域 Region Proposals

• Regions with Convolutional Neural Network
• 选出一些格子，在其上运行卷积网络
• segmentation algorithm 分割算法，选出色块
• 、