c++实现NMS

目录

What is NMS?

Relevant Concepts

Example Logic

Code

References​​​​​​​

What is NMS?

NMS - Non-maximum suppression, 非极大值抑制

在图像识别目标检测任务中，通常用框框（Bounding Box）表示识别出来的对象，对于同一个对象可能有很多个有着不同重叠区域的Bounding Box。每个Bounding Box对应的属于该特征的概率也不尽相同（confidence，置信度）。

所以NMS是用来抑制 非 最大置信度 的识别结果，只留下来置信度高的（是该特征的概率最大）结果。

Relevant Concepts

• Bounding Box
• IoU - Intersection over Union, 交并比 ： 两个Bounding Box的交集区域面积，除以并集区域面积

Example Logic

1. input-Bounding Box List和IoU的阈值， output-Bounding List with NMS processing
2. 按照confidence对Bounding Box从高到低排序
3. 计算Bounding Box两两的IoU, 如果IoU大于阈值，从List删除当前两个BBox中confidence的低的那个
4. 全部处理完，输出新的BBox List

Code

//NMS.H
#pragma once
void nms_test();

typedef struct BBox
{
	int x; //(x,y) left-up position, for x, left is increasing
	int y; // for y, down is increasing
	int w; //width
	int h; //hight
	float confidence;
}BBox;

//NMS_test.cpp
#include "NMS.h"
#include <vector>
#include <algorithm>
#include <iostream>

float IOU(BBox box1, BBox box2)
{
	float iou = 0;
	int x_over_l = std::max(box1.x, box2.x); //left x position of over area
	int x_over_r = std::min(box1.x + box1.w, box2.x + box2.w); //right position of over area
	int y_over_u = std::max(box1.y, box2.y);
	int y_over_d = std::min(box1.y + box1.h, box2.y + box2.h);
	int over_w = x_over_r - x_over_l;
	int over_h = y_over_d - y_over_u;
	if (over_w <= 0 || over_h <= 0)
	{
		iou = 0;
	}
	else {
		iou = (float)over_w * over_h / (box1.w * box1.h + box2.w * box2.h - over_w * over_h);
	}
	return iou;
}

std::vector<BBox> NMS(std::vector<BBox> &boxes, float threshold)
{
	std::vector<BBox>resluts;
	sort(boxes.begin(), boxes.end(), [](BBox a, BBox b) {return a.confidence > b.confidence;});
	while (boxes.size() > 0)
	{
		resluts.push_back(boxes[0]);
		int index = 1;
		while (index < boxes.size()) {
			float iou_value = IOU(boxes[0], boxes[index]);
			std::cout << "iou_value=" << iou_value << '\n';
			if (iou_value > threshold) {
				boxes.erase(boxes.begin() + index);
			}
			else {
				index++;
			}
		}
		boxes.erase(boxes.begin());
	}
	return  resluts;
}
void nms_test()
{
	std::cout << "========NMS test begin======" << '\n';
	std::vector<BBox> input;

	BBox box1 = { 10,10,10,10,0.5 };
	BBox box2 = { 0,0,20,20,0.6 };

	input.push_back(box1);
	input.push_back(box2);
	std::vector<BBox> res;
	res = NMS(input, 0.15);
	for (int i = 0; i < res.size(); i++) {
		printf("%d %d %d %d %f", res[i].x, res[i].y, res[i].w, res[i].h, res[i].confidence);
		std::cout << '\n';
	}
	std::cout << "========NMS test end======" << '\n';
}

//main.cpp
#include "NMS.h"
int main()
{
	nms_test();
}

References

NMS原理视频	【目标检测】指标介绍：BoundingBox、IOU、NMS_哔哩哔哩_bilibili
#pragma once	C/C++编程：#pragma once用法总结-CSDN博客
NMS code	c++实现nms算法 - 代码先锋网
NMS code	nms算法c++实现_核函数估计 c++-CSDN博客
sort	【C++STL】快速排序算法(sort)的原理与使用_std::sort_OpenC++的博客-CSDN博客
using	C++：using : using的四大用法总结-CSDN博客
vector	C++ vector的用法（整理）-CSDN博客
vector	C++中vector的初始化和at函数以及clear函数的使用_vector 初始化 clear_liebecl的博客-CSDN博客