Understand yolov1 Part II non maximum suppression (NMS) in prediction stage

introduction

1  Prediction stage , After the model has been trained , Start the prediction stage

2  NMS Non maximum suppression , My understanding of the role in this is , When the category is the same , Any box can be selected

More than one instance .

Occurring simultaneously than （iou）

Then cross and compare is ：

Non maximum suppression （NMS） 

In the last article 《 understand YOLOV1 Chapter one Prediction stage _code bean The blog of -CSDN Blog 》 We predicted 98 Candidate box , So what we need to do now is to screen out the most accurate ones .

NMS The first 1 round

First of all, we will introduce one of them （ Like dogs ） The candidate boxes of are sorted in descending order by full probability , Then compare all candidate boxes with the first one , If the intersection union ratio of candidate boxes is greater than a certain threshold , Just clear the probability corresponding to this candidate box （ It is equivalent to excluding the candidate box , Because the intersection ratio is too large, it means that it is a repeated choice , The probability is smaller than the maximum , So you can kill ）

As shown in the figure above , The intersection ratio of yellow box and green box is greater than 0.5 result , The green frame was killed .

NMS The first 2 round

  After the first round , Those who cross the first box and are too big have been killed , So we started the second round , At this time of the second round Start with the candidate box with the second largest probability , That is, the blue frame in the figure , At this time, the red frame behind the blue frame is too large , So the red frame was killed .

There may be a third N round , And then to the end .

But so far , The yellow and blue boxes are the selection boxes that are ultimately retained .

NMS Non maximum suppression , My understanding of the role in this is , When the category is the same , Any box can be selected

More than one instance . 

---

Reference resources ：

Tongji Zihao brother's personal space _ Bili, Bili _bilibili