forked from jinsheng124/yolox
-
Notifications
You must be signed in to change notification settings - Fork 0
/
test_my_poly.py
136 lines (135 loc) · 5.68 KB
/
test_my_poly.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
import numpy as np
import math
from shapely.geometry import Polygon
import torch.nn as nn
import torch
def xyxya2corner(p):
#batchsize*5
#旋转矩阵
bs = p.shape[0]
corners = np.zeros((bs,4,2))
for i in range(bs):
w = p[i][2] - p[i][0]
h = p[i][3] - p[i][1]
set_x = (p[i][2] + p[i][0])/2.0
set_y = (p[i][3] + p[i][1])/2.0
poly_set = np.array([[-w/2,-h/2],[w/2,-h/2],[w/2,h/2],[-w/2,h/2]])
a = p[i][4]
a_mu = np.array([[np.cos(a),np.sin(a)],[-np.sin(a),np.cos(a)]])
poly_set = poly_set@a_mu
poly_c = np.array([set_x,set_y])
poly_c = np.expand_dims(poly_c,0).repeat(4,axis=0)
corners[i] = poly_c + poly_set
return corners
def corner2xyxya(points):
w = ((points[0][0] - points[1][0]) ** 2 + (points[0][1] - points[1][1]) ** 2) ** 0.5
h = ((points[1][0] - points[2][0]) ** 2 + (points[1][1] - points[2][1]) ** 2) ** 0.5
cx = (points[0][0] + points[2][0]) / 2.0
cy = (points[0][1] + points[2][1]) / 2.0
if points[1][0] - points[0][0] == 0:
a = math.pi / 2.0
else:
a = math.atan((points[1][1] - points[0][1]) / (points[1][0] - points[0][0]))
keypoint = [int(cx - w/2.0),int(cy - h/2.0),int(cx+w/2.0),int(cy+h/2.0),a]
return np.array(keypoint)
def poly_iou(x1,x2):
x1 = np.asarray(x1)
x2 = np.asarray(x2)
x1 = Polygon(x1[:8].reshape((4,2)))
x2 = Polygon(x2[:8].reshape((4, 2)))
if not x1.is_valid or not x2.is_valid:
return 0
inter = Polygon(x1).intersection(Polygon(x2)).area
union = x1.area + x2.area - inter
if union==0:
return 0
else:
return inter/union
def gaussian_label(label, num_class, u=0, sig=4.0):
'''
转换成CSL Labels：
用高斯窗口函数根据角度θ的周期性赋予gt labels同样的周期性，使得损失函数在计算边界处时可以做到“差值很大但loss很小”；
并且使得其labels具有环形特征，能够反映各个θ之间的角度距离
@param label: 当前box的θ类别 shape(1)
@param num_class: θ类别数量=180
@param u: 高斯函数中的μ
@param sig: 高斯函数中的σ
@return: 高斯离散数组:将高斯函数的最高值设置在θ所在的位置，例如label为45，则将高斯分布数列向右移动直至x轴为45时，取值为1 shape(180)
'''
# floor()返回数字的下舍整数 ceil() 函数返回数字的上入整数 range(-90,90)
# 以num_class=180为例，生成从-90到89的数字整形list shape(180)
x = np.array(range(math.floor(-num_class / 2), math.ceil(num_class / 2), 1))
y_sig = np.exp(-(x - u) ** 2 / (2 * sig ** 2)) # shape(180) 为-90到89的经高斯公式计算后的数值
# 将高斯函数的最高值设置在θ所在的位置，例如label为45，则将高斯分布数列向右移动直至x轴为45时，取值为1
return np.concatenate([y_sig[math.ceil(num_class / 2) - int(label.item()):],
y_sig[:math.ceil(num_class / 2) - int(label.item())]], axis=0)
if __name__=="__main__":
# 转换格式
#将长边放在第三位，短边放在第4位，这样的角度更加符合直觉
# classid,x_c,y_c,longside,shortside
# po = [
# [
# 3532.7797967589277,
# 2830.1922922713475
# ],
# [
# 3513.3806027343203,
# 2798.0623771680916
# ],
# [
# 3542.7270651309564,
# 2780.3437583625755
# ],
# [
# 3562.1262591555637,
# 2812.4736734658313
# ]
# ]
# poo = [
#
# [
# 3513.3806027343203,
# 2798.0623771680916
# ],
# [
# 3542.7270651309564,
# 2780.3437583625755
# ],
# [
# 3562.1262591555637,
# 2812.4736734658313
# ],
# [
# 3532.7797967589277,
# 2830.1922922713475
# ]
# ]
# k = poly_iou(po,poo)
# print(k)
# po_1 = corner2xyxya(po)
# print(po_1)
# po_1 = np.expand_dims(po_1,axis=0)
# po_2 = xyxya2corner(po_1)
# print(po_2[0])
#角度损失计算
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
BCEangle = nn.BCEWithLogitsLoss()
#预测值
ps = torch.randn(4,180)
csl_label_flag = False
if not csl_label_flag:
#直接用交叉熵损失，即将角度分为180份，每一份都是一个预测值
ttheta = torch.full_like(ps, 0)
ttheta[range(4), 90] = 1
langle_loss = BCEangle(ps, ttheta) # BCE Θ类别损失以BCEWithLogitsLoss来计算
else:
ttheta = torch.zeros_like(ps) # size(num, 180)
for idx in range(len(ps)): # idx start from 0 to len(ps)-1
# 3个tensor组成的list (tensor_angle_list[i]) 对每个步长网络生成对应的class tensor tangle[i].shape=(num_i, 1)
theta = torch.tensor(90.) # 取出第i个layer中的第idx个目标的角度数值 例如取值θ=90
# CSL论文中窗口半径为6效果最佳，过小无法学到角度信息，过大则角度预测偏差加大
#此处用了高斯平滑
csl_label = gaussian_label(theta, 180, u=0, sig=6) # 用长度为1的θ值构建长度为180的csl_label
ttheta[idx] = torch.from_numpy(csl_label) # 将cls_label放入对应的目标中
langle_loss = BCEangle(ps, ttheta)
print(langle_loss)