三维点云处理06-2D/3DIoU计算

三维点云处理06-2D/3DIoU计算

• IoU定义：预测边框与真是边框的交集与并集之比
• 首先说明IoU的计算并不需要像下图所示对各种情况进行分类后，再进行计算
  
• IoU的计算本质上是集合交集的计算，首先从一维IoU的计算开始，然后扩展到二维等等
  

一维IoU计算

• 首先看上图右侧一维IoU的计算，IoU的计算本质就是集合上下边界的计算

    1.计算交集的下界：max(x1,y1)
    2.计算交集的下界：min(x2,y2)
    3.判断交集是否真是存在：上界-下界： min(x2,y2) - max(x1,y1)
    如果上界-下界 > 0     交集存在
    如果上界 - 下界 <=0 交集不存在
    4.计算并集：集合A + 集合B - 集合A和B的交集
    (x2 - x1) + (y2 - y1) - (min(x2,y2) - max(x1,y1))
    5.计算IoU
  

def iou(set_a, set_b):
'''
一维IoU的计算
'''
	# 获得集合A和B的边界 
	x1, x2 = set_a
	y1, y2 = set_b
	# 计算交集的上下界
	low = max(x1,y1)
	high - min(x2, y2)
	# 计算交集
	if high - low < 0:
		inter = 0
	else:
		inter = high - low
	# 计算并集
	union = (x2 -x1) + (y2 - y1) - inter
	# 计算IoU
	iou = inter / union
	return iou

二维IoU计算

• 从一维IoU的计算拓展到二维IoU的计算，就是计算X和Y两个方向上交集的上界和下界

• 二维bbox的输入一般有两种形式：

    一种是中心+宽高的形式：(x,y,w,h);
    一种是左上和右下两个角的形式：（x1,y1,x2,y2）
  

def iou(box1,box2):
	'''
	二维IoU的计算：
	注意图像坐标系的原点一般在左上角，x方向朝右，y方向朝下
	box的表示:[top, left, bottom, right]
	box的表示分别对应:[y1,x1,y2,x2]
	'''
	in_h = min(box1[2],box2[2]) - max(box1[0],box2[0])
	in_w = min(box1[3],box2[3]) - max(box1[1],box2[1])
	inter = 0 if in_h  < 0 or in_w < 0 else in_h * in_w
	union = (box1[2] - box1[0]) * (box1[3] - box1[1]) +  (box2[2] - box2[0]) * (box2[3] - box2[1])  - inter
	iou = inter / union
	return iou

def iou(box1,box2):
	'''
	二维Iou计算
	box表示形式：[x,y,w,h]
	'''
	x_max = box1[0] + box1[2] /2
	y_max = box1[1] + box1[3] /2
	x_min = box1[0] - box1[2] /2
	y_min = box1[1] - box1[3] / 2

	l_x_max = box2[0] + box2[2] /2
	l_y_max = box2[1] + box2[3] /2
	l_x_min = box2[0] - box2[2] /2
	l_y_min = box2[1] - box2[3] / 2	

	inter_x_max = min(x_max,l_x_max)
	inter_x_min = max(x_min,l_x_min)
	inter_y_max = min(y_max, l_y_max)
	inter_y_min = max(y_min,l_y_min)

	inter_w = inter_x_max - inter_x_min
	inter_h = inter_y_max - inter_y_min

	inter = 0 if inter_w < 0 or inter_h < 0 else inter_w * inter_h
	union = (box1[2] * box1[3]) + (box2[2] * box2[3]) - inter
	iou = inter / union
	return iou

三维IoU计算

• 3D IoU的计算与2D和1D类似，核心思想仍然是寻找各个维度上的交集

def  iou(box1,box2):
	'''
	3D IoU计算
	box表示形式：[x1,y1,z1,x2,y2,z2] 分别是两对角点的坐标
	'''
	in_w = min(box1[3],box2[3]) - max(box1[0],box2[0])
	in_l = min(box1[4],box2[4]) - max(box1[1],box2[1])
	in_h = min(box1[5],box2[5]) - max(box1[2],box2[2])

	inter = 0 if in_w < 0 or in_l < 0 or in_h < 0 else in_w * in_l * in_h
	union = (box1[3] - box1[0]) * (box1[4] - box1[1]) * (box1[5] - box1[2]) + (box2[3] - box2[0]) * (box2[4] - box2[1]) * (box2[5] - box2[2])  - inter
	iou = inter / union
	return iou

def iou(box1,box2):
	'''
	3D Iou计算
	box表示形式：[x,y,z,w,l,h]
	''' 
	x_max = box1[0] + box1[3] / 2
	y_max = box1[1] + box1[4] / 2
	z_max = box1[2] + box1[5] / 2
	x_min = box1[0] - box1[3] / 2
	y_min = box1[1] - box1[4] / 2
	z_min = box1[2] - box1[5] / 2	

	l_x_max = box2[0] + box2[3] / 2
	l_y_max = box2[1] + box2[4] / 2
	l_z_max = box2[2] + box2[5] / 2
	l_x_min = box2[0] - box2[3] / 2
	l_y_min = box2[1] - box2[4] / 2
	l_z_min = box2[2] - box2[5] / 2

	inter_x_max = min(x_max, l_x_max)
	inter_x_min = max(x_min, l_x_min)
	inter_y_max = min(y_max, l_y_max)
	inter_y_min = max(y_min, l_y_min)
	inter_z_max = min(z_max,l_z_max)
	inter_z_min = max(z_min,l_z_min)

	inter_w = inter_x_max - inter_x_min
	inter_l = inter_y_max - inter_y_min
	inter_h = inter_z_max - inter_z_min

	inter = 0 if inter_w < 0 or inter_l < 0 or inter_h  < 0 else inter_w * inter_l * inter_h
	area1 = (x_max - x_min) * (y_max - y_min) * (z_max - z_min)
	area2 = (l_x_max - l_x_min) * (l_y_max - l_y_min) * (l_z_max - l_z_min)
	union = area1 + area2 - inter
	iou = inter / union
	return iou

2维旋转IoU的计算

• 使用opencv库函数实现

def iou_rotate_calculate(boxes1, boxes2):
	'''
	2D旋转框的表示：[x ,y, w, h, theta],旋转角theta单位是角度，而不是弧度
	'''
	# 计算boxes1中旋转框的面积
	area1 = boxes1[:,2] * boxes1[:,3]
	# 计算boxes2中旋转框的面积
	area2 = boxes2[:,2] * boxes2[:,3]
	# 创建用于存储iou的list
	ious = []
for i, box1 in enumerate(boxes1):
	temp_ious = []
	r1 = ((box1[0],box1[1]),(box1[2],box1[3]),box1[4])
	for j, box2 in enumerate(boxes2):
		r2 = ((box2[0],box2[1]),(box2[2],box2[3]),box2[4])

		int_pts = cv2.rotatedRectangleIntersection(r1,r2)[1]
		if in_pts is not None:
			order_pts = cv2.convexHull(in_pts, returnPoints = True)
			int_area = cv2.contourArea(order_pts)

			iou = in_area * 1.0 / (area1[i] + area2[j] - int_area)
			temp_ious.append(iou)
		else:
			temp_ious.append(0.0)
	ious.append(temp_ious)
return np.array(ious,dtype=np.float32)