Understand the first prediction stage of yolov1

First of all to see YOLOV1 Network structure , Feeling is ordinary convolutional neural network ：

  The final output is a 7*7*30 The black box of , Think of it as a book 7*7 The size has 30 Page of a book .

  Every page of the book , from 49 A square （grid cell） form , V1 Each in the version grid cell forecast 2 Boxes （bounding box,bounding box The center of lies in this gridcell） So we can generate 98 individual bounding box.

Every bounding box, contain 5 Parameters ：

1  x,y  bounding box The location of the center point of

2 h,w  bounding box The width and height of

3 c Whether to include the confidence of the target object （ This confidence , It should be cross comparison ）

The perspective is switching back grid cell,grid cell It also includes the probability of each category , use grid cell The category probability of , multiply grid cell Self generated bounding box The degree of confidence , Get this bounding box Probability corresponding to each category .

In this case , Filter out those with low probability , The frame containing location information and category information can be drawn .

Look at the black box again ：

that , We just looked at the left view of the black box , That just mentioned every grid cell Also includes A lot of information . Then the front view of this picture will be clearer （ And each or 49 individual grid cell Are all made of depth , This depth is 30）.

here , Let's focus on , Actually, one. grid cell And the depth information it contains ：

As mentioned before , Every grid cell Predictable 2 Boxes （bounding box）, Purple and green are these two bounding box part , Every bounding box contain 5 Parameters , Yes, it's the one mentioned above 5 Parameters . Last 20 Dimensions store the grid cell Corresponding 20 Category probability . Seeing here, we probably know .V1 Version a picture can generate at most 98 Boxes , At most 20 Species category .

Summary

For better understanding , Here are a few points ：

1. grid cell and bounding box There is a corresponding relationship ： Every grid cell Corresponding to the two bounding box
2. bounding box Include the confidence level of whether the object is included （ Here we need to further confirm the confidence of what it is ）
3. grid cell Include the probability of each category .
4. Two bounding box Each has a confidence , If the figure below shows ：

  The red box represents grid cell, Two black frames represent grid cell Two generated bounding boxes, The thicker the line, the higher the confidence . Two bounding box Share the grid cell Probability of category . Multiply probability and confidence , Will get this category to judge Total probability .

Look at this dynamic diagram , Should be able to better understand ：

---

 

Continue analysis ： If each grid Apply your own color to the maximum probability of the middle category , Then you may get the following effects ：

Here is another conclusion ,V1 At most versions are predicted 20 Species , And at most, you can box out 49 An object . 

---

Reference material

【 intensive reading AI The paper 】YOLO V1 object detection , Just look at me _ Bili, Bili _bilibili