1. 规则矩形框的IoU计算
有些目标检测中,预测的边界框为规则的矩形,则只需要知道矩形的左上角和右下角的坐标信息,就可以得到矩形框所有想要的信息。对于这种情况,IoU的python实现如下(python3.5)
def IoU(box1, box2):
'''
计算两个矩形框的交并比
:param box1: list,第一个矩形框的左上角和右下角坐标
:param box2: list,第二个矩形框的左上角和右下角坐标
:return: 两个矩形框的交并比iou
'''
x1 = max(box1[0], box2[0]) # 交集左上角x
x2 = min(box1[2], box2[2]) # 交集右下角x
y1 = max(box1[1], box2[1]) # 交集左上角y
y2 = min(box1[3], box2[3]) # 交集右下角y
overlap = max(0., x2-x1) * max(0., y2-y1)
union = (box1[2]-box1[0]) * (box1[3]-box1[1]) \
+ (box2[2]-box2[0]) * (box2[3]-box2[1]) \
- overlap
return overlap/union
if __name__ == '__main__':
# box = [左上角x1,左上角y1,右下角x2,右下角y2]
box1 = [10, 0, 15, 10]
box2 = [12, 5, 20, 15]
iou = IoU(box1, box2)
2. 非矩形框IoU计算
在有些目标检测中,检测框并不是规则的矩形框,例如自然场景下的文本检测,有些呈现平行四边形,梯形等情况,这时计算IoU时,就比较复杂一些。这时可以借助于python的一些库实现多边形的面积计算。
import shapely
import numpy as np
from shapely.geometry import Polygon, MultiPoint # 多边形
def bbox_iou_eval(box1, box2):
'''
利用python的库函数实现非矩形的IoU计算
:param box1: list,检测框的四个坐标[x1,y1,x2,y2,x3,y3,x4,y4]
:param box2: lsit,检测框的四个坐标[x1,y1,x2,y2,x3,y3,x4,y4]
:return: IoU
'''
box1 = np.array(box1).reshape(4, 2) # 四边形二维坐标表示
# python四边形对象,会自动计算四个点,并将四个点重新排列成
# 左上,左下,右下,右上,左上(没错左上排了两遍)
poly1 = Polygon(box1).convex_hull
box2 = np.array(box2).reshape(4, 2)
poly2 = Polygon(box2).convex_hull
if not poly1.intersects(poly2): # 如果两四边形不相交
iou = 0
else:
try:
inter_area = poly1.intersection(poly2).area # 相交面积
iou = float(inter_area) / (poly1.area + poly2.area - inter_area)
except shapely.geos.TopologicalError:
print('shapely.geos.TopologicalError occured, iou set to 0')
iou = 0
return iou
if __name__ == '__main__':
# box = [四个点的坐标,顺序无所谓]
box3 = [10, 0, 15, 0, 15, 10, 10, 10] # 左上,右上,右下,左下
box4 = [12, 5, 20, 2, 20, 15, 12, 15]
iou = bbox_iou_eval(box3, box4)
print(iou)
References
IOU交并比的计算