week03.Ng's CNN Course

CNN 2017.12

Detection Algorithms

1. Object Localization

• What are localization and detection？

  • Image Classification
  • classification with localization
  • Detection

• Defining the target label $y$

  1. pedestrian
  2. car
  3. motorcycle

  4. background

     • 定义label相当于定义最后网络的输出

2. Landmark Detection

• landmark detection

  • e.g.人脸的landmark检测，可以设计输出如下

    一个output代表存在人脸的概率
    人脸的landmark($l_{1x},l_{1y},l_{2x},l_{2y},l_{3x},l_{3y},l_{4x},l_{4y}\ldots,l_{64x},l_{64y}$)

  • pose detection

    key points in the person's body

3. Object Detection

sliding windows

• car detection example

  先利用一些小图片，只含有车或不含有，来训练一个分类器。
  然后利用sliding windows detection。

• sliding windows detection

  用不同大小的windows在图像中滑动，判断该小块区域是否是要检测的目标。

4. Convolutional Implementation of Sliding Windows

[Sermanet et al, 2014, OverFeat: Integrated recognition, localization and detection using convolutional networls]

• Turning FC layer into convolutional layers

  Kernel的大小和input volume的宽高一样。

• Convolution Implementation of sliding windows

  

  • 先训练一个针对特定大小图片的classfier，全连接层用卷积来做，这样可以接受任意尺寸大小的图像输入，只是最后输出的个数不同。

5. Bounding Box Predictions

• Output accurate bounding boxes

• YOLO algorithm

  [Redmon et al, 2015, You Only Loo Once: Unified real-time object detection]

  • place a grid on the image 将图像分成几个栅格

    e.g. 将100x100的图像分成3x3的9个格子

  • Labels for training

    $$\begin{align*} y = \left[p_c \ b_x\ b_y \ b_h \ b_w \ c_1 \ c_2 \ c_3 \right] \end{align*}$$

    最后网络的输出就是3x3x8的格式。

  • 中心$\left[b_x \ b_y\right]$在哪个栅格内，就判定object在哪个栅格

  • 可以采用更多的栅格划分使得objects的中心点落在同一个栅格内的概率更小

• Specify the bounding boxes

6. Intersection Over Union(IOU)

• Evaluating object localization

  • IOU = ( size of intersection) / (size of union)

    两个box的合集叫union，intersection就是二者的交集。
    Usually, the answer is right if $IOU>0.5$。

  • question

    要是groundtruth box 围住了predicted box呢？

7. Non-max Suppression

• you may find multiple detections of the same objects

• Non-max supperssion example

  clean up the redundant detections

  选择一个confidence最大的bounding box然后去除掉与其的IOU大于一定阈值的bounding box，最后剩下的就是检测到的最终结果。

  

8. Anchor Boxes

• One grid cell can only detect one object

• overlapping objects

  predefined different shapes called anchor boxes

• Anchor box algorithm

9. YOLO algorithm

• Training

• Prediction

• NMS

10. Region proposal: R-CNN

• Regions with CNN

  • segmentation algorithm
    选择一些较为合理的区域。

  • a little bit slow