Recorded target detection NMS (non maximum suppression)

NMS thought ：

1. First, sort all boxes by confidence , Select the box with the highest confidence , Write it down as S

2. Remaining boxes and S do IOU Calculation , Greater than the set threshold threshold （ permanent 0.7） Others are suppressed , The most basic suppression method is to set the confidence of these boxes 0.

3. Continue to find the one with the highest score in the remaining boxes , repeat 1,2. To the box without repetition , Algorithm stop .

pytorch Realization （ quote ：https://blog.csdn.net/KANG157/article/details/124649838）：

For specific analysis, I gave comments in the code

def NMS(boxes,scores, thresholds):
    x1 = boxes[:,0]      # Top left x
    y1 = boxes[:,1]      # Top left y
    x2 = boxes[:,2]      # Lower right x
    y2 = boxes[:,3]      # Lower right y
    areas = (x2-x1)*(y2-y1)     #  Calculate the area of each box , There is aeras in 
 
    _,order = scores.sort(0,descending=True)   #  Rank the scores of all boxes in descending order , return order Indexes 
    keep = []   
    while order.numel() > 0:                   #  When order The number of elements is not 0
        i = order[0]                           #  Get the index with the highest score and assign it to i, Remember this box as S
        keep.append(i)                         #  Add in keep Set 
        if order.numel() == 1:                 #  If there is only one box , Then the algorithm ends 
            break 
        xx1 = x1[order[1:]].clamp(min=x1[i])   #  adopt clamp function , Check the remaining boxes and S The upper left point of the intersection x value 
        yy1 = y1[order[1:]].clamp(min=y1[i])   #  adopt clamp function , Check the remaining boxes and S The upper left point of the intersection y value 
        xx2 = x2[order[1:]].clamp(max=x2[i])   #  adopt clamp function , Check the remaining boxes and S The lower right point of the intersection x value 
        yy2 = y2[order[1:]].clamp(max=y2[i])   #  adopt clamp function , Check the remaining boxes and S The upper left point of the intersection x value 
 
        w = (xx2-xx1).clamp(min=0)  # Will intersect width Negative values are limited to 0, The goal is a box without intersection , The intersection part is set to 0
        h = (yy2-yy1).clamp(min=0)  # Will intersect height Negative values are limited to 0, ditto 
        inter = w*h #  Calculate the intersection area 
 
        ovr = inter/(areas[i] + areas[order[1:]] - inter)  # Calculation iou
        ids = (ovr<=thresholds).nonzero().squeeze() # Get all iou The index of the box whose value is less than the threshold 
        if ids.numel() == 0:  # When does not exist and the current box iou When the value is less than the threshold box , Code end 
            break 
        order = order[ids+1]  # Pay attention to the addition here 1, because ids The calculation is in addition to the first box , Make a recovery to order Step back 1 individual , So add 1
    return torch.LongTensor(keep)

Function analysis ：

torch.sort(input, dim=0, descending=False)

dim： Sort dimension 
descending： Descending or not 
return：1. An ordered array 
        2. The corresponding index  

torch.clamp(input, min=a, max=b)

usage： take input Data is limited to [a,b] In closed interval 
 principle ：
input[i] = a,           if input[i] <= a
         = input[i]     if input[i] > a and input[i] < b
         = b            if input[i] >= b

torch.nonzero(input)

return： input Central Africa 0 Index of elements