NMS代码(python,C++)

NMS流程代码

IOU计算
- python代码
- C++代码
NMS计算代码
opencv（c++）显示NMS结果

IOU计算

python代码

import numpy as np

area1 = np.array(([1,1],[1,3],[3,3],[3,1]))
area2 = np.array(([2,0],[4,0],[4,2],[2,2]))

xy1_min = np.min(area1,axis=0)
xy1_max = np.max(area1,axis=0)
xy2_min = np.min(area2,axis=0)
xy2_max = np.max(area2,axis=0)


Ux_min = max(xy1_min[0],xy2_min[0])
Ux_max = min(xy1_max[0],xy2_max[0])
Uy_min = max(xy1_min[1],xy2_min[1])
Uy_max = min(xy1_max[1],xy2_max[1])

#交集面积
area1 = (Ux_max - Ux_min) * Uy_max - Uy_min

#并集面积
area2 = (xy1_max[0] - xy1_min[0]) * (xy1_max[1] - xy1_min[1])
area3 = (xy2_max[0] - xy2_min[0]) * (xy2_max[1] - xy2_min[1]) 
res = area1 / (area2 + area3 - area1)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23

C++代码

#include
#include
#include 
#include
using namespace std;

vector<int> get_mins(vector<vector<int>>& Arr)
{
	vector<int> res(Arr[0].size(), INT_MAX);
	for (int i = 0; i < Arr.size(); i++)
	{
		for (int j = 0; j < Arr[i].size(); j++)
		{
			if (res[j] > Arr[i][j])
				res[j] = Arr[i][j];
		}
	}
	return res;
}

vector<int> get_maxs(vector<vector<int>>& Arr)
{
	vector<int> res(Arr[0].size(), INT_MIN);
	for (int i = 0; i < Arr.size(); i++)
	{
		for (int j = 0; j < Arr[i].size(); j++)
		{
			if (res[j] < Arr[i][j])
				res[j] = Arr[i][j];
		}
	}
	return res;
}


int main()
{
	vector<vector<int>> Arr1 = { {1,1},{3,3},{1,3},{3,1} };
	vector<vector<int>> Arr2 = { {2,0},{4,0},{2,2},{4,2} };

	vector<int> xy1_min = get_mins(Arr1);
	vector<int> xy1_max = get_maxs(Arr1);
	vector<int> xy2_min = get_mins(Arr2);
	vector<int> xy2_max = get_maxs(Arr2);
	//求交集
	int x_min = max(xy1_min[0], xy2_min[0]);
	int x_max = min(xy1_max[0], xy2_max[0]);
	int y_min = max(xy1_min[1], xy2_min[1]);
	int y_max = min(xy1_max[1], xy2_max[1]);
	int area1 = (x_max - x_min) * (y_max - y_min);

	cout << area1 << endl;
	//求并集
	int area2 = (xy1_max[0] - xy1_min[0]) * (xy1_max[1] - xy1_min[1]);
	int area3 = (xy2_max[0] - xy2_min[0]) * (xy2_max[1] - xy2_min[1]);

	float  res = float(area1) / (area3 + area2 - area1);
	cout << res << endl;
	return 0;
}

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

NMS计算代码

#include
#include
#include 

using namespace std;

float compute_iou(vector<float>& Arr1, vector<float>& Arr2)
{
	float h = min(Arr1[1], Arr2[1]) - max(Arr1[0], Arr2[0]);
	float w = min(Arr1[3], Arr2[3]) - max(Arr1[2], Arr2[2]);
	float area1 = h * w;

	float area2 = (Arr1[1] - Arr1[0]) * (Arr1[3] - Arr1[2]);
	float area3 = (Arr2[1] - Arr2[0]) * (Arr2[3] - Arr2[2]);
	return area1 / (area2 + area2 - area2);
}

vector<vector<float>> compute_nms(vector<vector<float>>& infos, float iou_theta, float theta_score)
{
	vector<vector<float>> filter_infos;
	//过滤分数达标的box
	for (auto info : infos)
	{
		if (info[4] >= theta_score)
			filter_infos.push_back(info);
	}

	vector<vector<float>> res;
	vector<bool> set_idx(filter_infos.size(),true);//为使用的info索引

	while (true)
	{
		//取score最大的box
		int temp_idx = 0;
		bool flag = false;//记录是否有为处理的box
		
		for (int i = 1; i < filter_infos.size() && set_idx[i]; i++)
		{
			if (filter_infos[i][4] > filter_infos[temp_idx][4])
			{
				flag = true;
				temp_idx = i;
			}
		}

		if (!flag)	break;
		set_idx[temp_idx] = false;
		vector<float> temp = filter_infos[temp_idx];
		res.push_back(temp);
		//计算选出的score最大的box于其他未使用的box的iou
		vector<int> iou_idx;
		for (int i = 0; i < filter_infos.size() & set_idx[i]; i++)
		{
			if (compute_iou(temp, filter_infos[i]) >= iou_theta)
			{
				res.push_back(filter_infos[i]);
				set_idx[i] = false;
			}
		}
	}
	return res;	
}

int main()
{
	vector<vector<float>> infos = { {30, 10, 200, 200, 0.95},
									{25, 15, 180, 220, 0.98},
									{35, 40, 190, 170, 0.96},
									{60, 60, 90, 90, 0.3},
									{20, 30, 40, 50, 0.1} };

	vector<vector<float>> res = compute_nms(infos, 0.5, 0.5);
	for (int i = 0; i < res.size(); i++)
	{
		for (int j = 0; j < res[0].size(); j++)
		{
			cout << res[i][j] << " ";
		}
		cout << endl;
	}
	return 0;
}
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

opencv（c++）显示NMS结果

在这里插入图片描述

code

#include
#include
#include 
#include
#include "stdc++.h"
using namespace cv;
using namespace std;

float compute_iou(vector<float>& Arr1, vector<float>& Arr2)
{
	//[x,y,w,h,score]
	//[x-w/2.y+h/2] [x+w/2]
	//求交集
	float x_min = max(Arr1[0] - Arr1[2] / 2, Arr2[0] - Arr2[2]/2);
	float x_max = min(Arr1[0] + Arr1[2] / 2, Arr2[0] + Arr2[2] / 2);
	float y_min = max(Arr1[1] - Arr1[3] / 2, Arr2[1] - Arr2[3] / 2);
	float y_max = min(Arr1[1] + Arr1[3] / 2, Arr2[1] + Arr2[3] / 2);
	float area1 = (x_max - x_min) * (y_max - y_min);

	//求并集
	float area2 = Arr1[2] * Arr1[3];
	float area3 = Arr2[2] * Arr2[3];
	float res = area1 / (area2 + area3 - area1);
	return res;
}

vector<vector<float>> compute_nms(vector<vector<float>>& infos, float iou_theta, float theta_score)
{
	vector<vector<float>> filter_infos;
	//过滤分数达标的box
	for (auto info : infos)
	{
		if (info[4] >= theta_score)
			filter_infos.push_back(info);
	}

	vector<vector<float>> res;
	vector<bool> set_idx(filter_infos.size(),true);//为使用的info索引

	while (true)
	{
		//取score最大的box
		int temp_idx = 0;
		bool flag = false;//记录是否有为处理的box
		int i = 1;
		for (;i < filter_infos.size(); i++)
		{
			if (set_idx[i] && filter_infos[i][4] > filter_infos[temp_idx][4])
			{
				flag = true;
				temp_idx = i;
			}
		}
		if (!flag)	break;
		set_idx[temp_idx] = false;
		vector<float> temp = filter_infos[temp_idx];
		res.push_back(temp);
		//计算选出的score最大的box于其他未使用的box的iou
		vector<int> iou_idx;
		for (int j = 0; j < filter_infos.size(); j++)
		{
			if (set_idx[j])
			{
				if (compute_iou(temp, filter_infos[j]) >= iou_theta)
				{
					//res.push_back(filter_infos[i]);
					set_idx[j] = false;
				}
			}

		}
	}

	return res;	
}

void show_res(cv::Mat src, vector<vector<float>> infos,vector<vector<float>> colors)
{
	src.depth();
	for (int i = 0; i < infos.size(); i++)
	{
		cv::Point P0 = cv::Point(int(infos[i][0] - infos[i][2] / 2), int(infos[i][1] + infos[i][3] / 2));
		cv::Point P1 = cv::Point(int(infos[i][0] + infos[i][2] / 2), int(infos[i][1] - infos[i][3] / 2));
		cv::Scalar color = cv::Scalar(colors[i][0], colors[i][1], colors[i][2]);
		rectangle(src, P0, P1, color, 2, 8, 0);
	}
	cv::imshow("Lena", src);
	cv::waitKey(0);
	cv::destroyAllWindows();
	return;
}

int main()
{
	string IMG_PATH = "C:/Users/hjfen/Pictures/Saved Pictures/";
	vector<vector<float>> infos = { {330, 310, 300, 200, 0.95},
									{325, 315, 180, 220, 0.98},
									{335, 340, 190, 170, 0.96},
									{360, 360, 90, 90, 0.3},
									{320, 330, 40, 50, 0.1} };


	vector<vector<float>> new_infos = compute_nms(infos, 0.5, 0.5);

	vector<vector<float>> colors = { {125,125,0},
										{255,0,0},
										{0,255,0},
										{0,125,125},
										{100,100,100} };

	cout << "Built with OpenCV " << CV_VERSION << endl;
	cv::Mat src = imread(IMG_PATH + "2.jpg");
	cv::Mat src1 = imread(IMG_PATH + "2.jpg");
	show_res(src, infos, colors);

	show_res(src1, new_infos, colors);
	return 0;
}

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

相关阅读:
Docker磁盘空间不足如何解决
 递归算法深入解析
 2 springMVC-处理器方法的返回值ModeVeiw,String，void,Object,List<Object>,String对象
 智能化水库监控，水库雨情在线监控系统解决方案
 信号与槽的连接方式
 服务器感染了.Malloxx勒索病毒，如何确保数据文件完整恢复？
HazelEngine 学习记录 - Single Shader 2D Renderer
云计算 - 阿里云最佳云上实践介绍卓越架构
 设计模式——（装饰者模式）（组合模式）
HTTPS的工作流程
原文地址：https://blog.csdn.net/qhu1600417010/article/details/126274847