LFW精确度验证__c++双线程读写txt

此篇内容承接上篇：LFW精确度验证__python读写txt

参考博客：

void类型和void* 的用法

C++ 多线程

我在百度，知乎，github上搜了下lfw验证准确率的方法，资源很少，所以我将我感觉有用的链接都整理到了下面：

Labeled Faces in the Wild 官方网站：http://vis-www.cs.umass.edu/lfw/#views

pairs.txt 官方介绍：http://vis-www.cs.umass.edu/lfw/README.txt

LFW介绍整理：https://blog.csdn.net/jobbofhe/article/details/79416661

基于VGG-Face的人脸识别测试：https://blog.csdn.net/u013078356/article/details/60955197   【主要参考】

LFW的正确率,使用方法：https://blog.csdn.net/u014696921/article/details/70161852   【主要参考】

LFW_API：https://github.com/jakezhaojb/LFW_API

 

具体参考工程：https://github.com/KangKangLoveCat/insightface_ncnn

#include <vector>
#include <iostream>
#include <opencv2/opencv.hpp>
#include "arcface.h"
#include "mtcnn.h"

#include <algorithm>
#include <cmath>
#include <fstream>
#include <pthread.h>

using namespace cv;
using namespace std;

#define PAIR_LINES 6000

vector< pair<string, string> > pos_vec;
vector< pair<string, string> > neg_vec;

MtcnnDetector detector("../models");
Arcface arc("../models");

cv::Mat ncnn2cv(ncnn::Mat img)
{
    unsigned char pix[img.h * img.w * 3];
    img.to_pixels(pix, ncnn::Mat::PIXEL_BGR);
    cv::Mat cv_img(img.h, img.w, CV_8UC3);
    for (int i = 0; i < cv_img.rows; i++)
    {
        for (int j = 0; j < cv_img.cols; j++)
        {
            cv_img.at<cv::Vec3b>(i,j)[0] = pix[3 * (i * cv_img.cols + j)];
            cv_img.at<cv::Vec3b>(i,j)[1] = pix[3 * (i * cv_img.cols + j) + 1];
            cv_img.at<cv::Vec3b>(i,j)[2] = pix[3 * (i * cv_img.cols + j) + 2];
        }
    }
    return cv_img;
}

void preprocess(){
    fstream label("./label.txt");
    
    string left, right;
    int num;

    for(int i=0; i<PAIR_LINES; i++){
        label >> left >> right >> num;
        if(num == 1)    
            pos_vec.push_back(make_pair(left, right));
        else
            neg_vec.push_back(make_pair(left, right));
    }
}

void* pos(void* args){
    fstream pos("./pos_scores.txt", fstream::in | fstream::out | fstream::trunc);
    
    string pos_left, pos_right, neg_left, neg_right;
    int pair_id = 0;
    int num=0;
    Mat img1, img2;

    ncnn::Mat ncnn_img1, ncnn_img2, det1, det2;
    vector<FaceInfo> results1, results2;
    vector<float> feature1, feature2;

    for(auto pos_it = pos_vec.begin(); pos_it != pos_vec.end(); ++pos_it){
        pair_id++;
        pos_left = pos_it->first;
        pos_right = pos_it->second;
        img1 = imread(pos_left);
        img2 = imread(pos_right);
        ncnn_img1 = ncnn::Mat::from_pixels(img1.data, ncnn::Mat::PIXEL_BGR, img1.cols, img1.rows);
        ncnn_img2 = ncnn::Mat::from_pixels(img2.data, ncnn::Mat::PIXEL_BGR, img2.cols, img2.rows);
        results1 = detector.Detect(ncnn_img1);
        results2 = detector.Detect(ncnn_img2);
        if(results1.size()==0 || results2.size()==0) continue;
        det1 = preprocess(ncnn_img1, results1[0]);
        det2 = preprocess(ncnn_img2, results2[0]);
        feature1 = arc.getFeature(det1);
        feature2 = arc.getFeature(det2);
        float cal = calcSimilar(feature1, feature2);
        if(cal<0.35) ++num;
        pos << pair_id <<"\t"<< cal << "\t" <<pos_left<<"\t"<<pos_right<<endl;
        cout << pair_id << "pos finish" <<endl;
    }
    pos << pair_id << "\t" << num << endl;
    pos.close();
}

void* neg(void* args){
    fstream neg("./neg_scores.txt", fstream::in | fstream::out | fstream::trunc);
    
    string pos_left, pos_right, neg_left, neg_right;
    int pair_id = 0;
    int num=0;
    Mat img1, img2;

    ncnn::Mat ncnn_img1, ncnn_img2, det1, det2;
    vector<FaceInfo> results1, results2;
    vector<float> feature1, feature2;

    for(auto neg_it = neg_vec.begin(); neg_it != neg_vec.end(); ++neg_it){
        pair_id++;
        neg_left = neg_it->first;
        neg_right = neg_it->second;
        img1 = imread(neg_left);
        img2 = imread(neg_right);
        ncnn_img1 = ncnn::Mat::from_pixels(img1.data, ncnn::Mat::PIXEL_BGR, img1.cols, img1.rows);
        ncnn_img2 = ncnn::Mat::from_pixels(img2.data, ncnn::Mat::PIXEL_BGR, img2.cols, img2.rows);
        results1 = detector.Detect(ncnn_img1);
        results2 = detector.Detect(ncnn_img2);
        if(results1.size()==0 || results2.size()==0) continue;
        det1 = preprocess(ncnn_img1, results1[0]);
        det2 = preprocess(ncnn_img2, results2[0]);
        feature1 = arc.getFeature(det1);
        feature2 = arc.getFeature(det2);
        float cal = calcSimilar(feature1, feature2);
        if(cal>0.35) ++num;
        neg << pair_id <<"\t"<< cal << "\t" <<neg_left<<"\t"<<neg_right<<endl;
        cout << pair_id << "neg finish" <<endl;
    }
    neg << pair_id << "\t" << num << endl;
    neg.close();
}

int main(int argc, char* argv[])
{
    preprocess();
    pthread_t tids[2];
    int ret1 = pthread_create(&tids[0], NULL, pos, NULL);
    if(ret1 != 0)
        cout<<"pthread_create_error: error_code="<< ret1<<endl;
    int ret2 = pthread_create(&tids[1], NULL, neg, NULL);
    if(ret2 != 0)
        cout<<"pthread_create_error: error_code="<< ret2<<endl; 
    pthread_exit(NULL);
    return 0;
}

可以看到双线程的效果

一共6000组图片，3000组同人，3000组不同人，考虑到图片的质量不是很高（本身就是通过网络收集，像素也低）

加上图片有少量采集错误（不同人的照片错归为了同一人，影响测试结果的大概不到10组），所以我设置了一个比较低的阈值0.35进行测试

 

pos_scores.txt

3000组正样本中，有51组相似度低于0.35

neg_scores.txt

3000组负样本中，有0组相似度高于0.35，我肉眼搜索了一下最高0.30

总的来看6000组在阈值0.35的精确度为99.15，阈值设为0.3应该会有更好的表现

由于没有找到比较可靠的资料，所以测试方案是我自己设计的，虽然感觉这种测试精确度的方法不太科学，但相似度并没有一个准确的衡量方法

此方案算是1：1验证，也可以当作1：n来测试，但计算量会大些