计算IoU(D2L)

#IoU
#给定两个边界框或锚框的列表

def box_iou(boxes1,boxes2):
    """ 计算两个锚框或边界框列表中成对的交并比"""
    box_area = lambda boxes : ((boxes[:,2]-boxes[:,0])*(boxes[:,3]-boxes[:,1])) #box面积
    # boxes1:[nums_boxes1，4]
    # boxes2：[nums_boxes2,4]
    # areas1:[nums_boxes1,]
    # areas2:[nums_boxes2,]
    areas1 = box_area(boxes1)
    areas2 = box_area(boxes2)
    """ areas1:tensor([100., 225.]) areas2:tensor([144., 676.]) """
    #print(f"areas1:{areas1}\n areas2:{areas2}")
    """ inter_upperlefts ,inter_lowerrights,inters : shape:(nums_boxes1,nums_boxes2,2) box_axis:[xmin,ymin,xmax,ymax] """
    inter_upperlefts  = torch.max(boxes1[:,None,:2],boxes2[:,:2]) #[:2]左上角，两两组合所有左上角坐标max值
    inter_lowerrights = torch.min(boxes1[:,None,2:],boxes2[:,2:]) #[2:]右下角
    inters  =(inter_lowerrights - inter_upperlefts).clamp(min=0) #clamp确保不出现负数
    """ inter_upperlefts: 4组坐标左上角max值(与右下角一一对应) tensor([[[13., 13.], [22., 22.]], [[15., 15.], [22., 22.]]]) inter_lowerrights:4组坐标右下角min值 tensor([[[20., 20.], [20., 20.]], [[25., 25.], [30., 30.]]]) inters:(右下角-左上角)得到每一对box交集的(宽,高) tensor([[[ 7., 7.], [ 0., 0.]], [[10., 10.], [ 8., 8.]]]) """

    print(f"inters:{
      inters}")
    # inter_areas , union_areas ,shape:[nums_boxes1,nums_boxes2]
    inter_areas  =inters[:,:,0] * inters[:,:,1] #分子 每一组数据的宽和高相乘获得交集面积
    union_areas  =areas1[:,None] + areas2 - inter_areas #分母
    """ inter_areas: tensor([[[ 7., 7.], [ 0., 0.]], [[10., 10.], [ 8., 8.]]]) union_areas: tensor([[195., 776.], [269., 837.]]) """
    return inter_areas / union_areas
    
#-----Test-------------------- 
#假设每个box中有两组数据:[2,4]

boxes1 = [ [10.,10.,20.,20.],
           [15.,15.,30.,30.]]
boxes2 = [[13.,13.,25.,25.],
          [22.,22.,48.,48.]]
boxes1 = torch.tensor(boxes1)
boxes2 = torch.tensor(boxes2)
print(boxes2.shape)
box_iou(boxes1,boxes2)


    



    

输出：

tensor([[0.25, 0.00],
        [0.37, 0.08]])

如何理解左上角和右下角坐标计算代码？

x =torch.arange(4,8).view(2,2)
y = torch.arange(4).view(2,2)
print(x,y)
""" x:tensor([[4, 5], [6, 7]]) ---------------------- y:tensor([[0, 1], [2, 3]]) ------------------------------------- x[:,None,:]: 根据广播机制，x会在中间None创建的新维度复制数据形成[2,2,2]以匹配y的维度[2,2]，只有在计算的时候才会复制，打印不出来 tensor([[[4, 5], [4,5]], [[6, 7], [6,7]]]) """
print(x[:,None,:]-y[:,:])  #shape:[2,1,2]-[2,2]
""" 输出： tensor([[[4, 4], [2, 2]], [[6, 6], [4, 4]]]) """