mobilenetssd 版caffe实现

lijingle

在讲解本文时首先看下效果，使用的数据量不多，并且训练并非标准尺寸

object_detect

这里环境使用的是ubuntu16.04，如果使用更高的版本，也是可以的但是caffe的一些配置文件需要改动，在训练之前首先要安装caffe，本文采用的是原版caffe 对原版caffe进行编译安装。下载mobilenetssd源码以及相关文件，这里我放到了百度网盘。

链接：https://pan.baidu.com/s/1rJhQu-uVjVU9HUIf-Hg5Vg 

提取码：ps9b

下载后进行解压，然后将代码放到SSD_HOME/examples/下如图：


制作数据集：

这里可以参考文章 进行标注数据，标注如图：


其中一个为图片，一个为xml文件。

创建数据集

#进到代码文件下

```cd create_lmdb/code```

修改labelmap.prototxt文件，添加自己标注的图片的类，例如添加cat，dog等

修改create_list.sh 和 create_data.sh文件里的路径，分别运行脚本

./create_list.sh

./bash create_data.sh

这样就会在文件夹中出现trainval.txt, test.txt, test_name_size.txt文件。


训练数据
第四课

object_detect

这里在example/MobileNetSSD_train.prototxt文件中修改你的lmdb文件路径，就可以进行训练，如图所示：


执行./train.sh就进行训练模型

如果想运行较快，还可以进行去除bn层，这里直接运行python merge_bn.py --model example/MobileNetSSD_deploy.prototxt --weights snapshot/mobilenet_iter_xxxxxx.caffemodel就可以得到相关的deploy文件和模型文件。
上述视频是根据c++进行预测的下面为相关代码，可以根据需求自行修改(opencv4.5)：

#include <stdio.h>

#include <opencv2/opencv.hpp>

#include <opencv2/dnn.hpp>

#include <opencv2/highgui.hpp>

#include <iostream>

#include <opencv2/core/ocl.hpp>

#include <iostream>

using namespace cv;

using namespace std;

const size_t width = 500;

const size_t height = 120;

size_t det_index;

const float scaleFector = 0.007843f;

const float meanVal = 127.5;

dnn::Net net;

const char* class_video_Names[] = { "background",

"yiwu"};

Mat detect_from_video(Mat &src,float &x1,float &y1,float &x2,float &y2,size_t &det_index)

{

  Mat blobimg = dnn::blobFromImage(src, scaleFector, Size(width, height), Scalar(127.5, 127.5, 127.5));

  net.setInput(blobimg, "data");

  Mat detection = net.forward("detection_out");

  //cout << detection.size[2]<<" "<< detection.size[3] << endl;

  Mat detectionMat(detection.size[2], detection.size[3], CV_32F, detection.ptr<float>());

  const float confidence_threshold = 0.75;

  for(int i=0; i<detectionMat.rows; i++){

  float detect_confidence = detectionMat.at<float>(i, 2);

  if(detect_confidence > confidence_threshold){

    det_index = (size_t)detectionMat.at<float>(i, 1);

    x1 = detectionMat.at<float>(i, 3)*src.cols;

    y1 = detectionMat.at<float>(i, 4)*src.rows;

    x2 = detectionMat.at<float>(i, 5)*src.cols;

    y2 = detectionMat.at<float>(i, 6)*src.rows;

  }

  }

  return src;

}



Mat changeImgLeft(Mat img1,Mat dst)

{

    

    Mat roi;    

    cout<<img1.rows<<img1.cols<<endl;

    roi = Mat::zeros(img1.size(),CV_8U);

    imshow("roi", roi);

    vector<vector<Point>> contour_left;

    vector<Point> pts_left;

    pts_left.push_back(Point(640, 281));

    pts_left.push_back(Point(246, 257));

    pts_left.push_back(Point(244, 247));

    pts_left.push_back(Point(97, 237));

    pts_left.push_back(Point(83, 168));

    pts_left.push_back(Point(640, 160));

    contour_left.push_back(pts_left);

    drawContours(roi,contour_left,0,Scalar::all(255),-1);

    imshow("roi_d", roi);

    img1.copyTo(dst,roi);

    imshow("dst", dst);

    

    return dst;

}



int main(int argc,char ** argv)

{

  float f,x1=0.0,y1=0.0,x2=0.0,y2=0.0;

  float FPS[16];

  int i, Fcnt=0;

  int count =0;

  VideoWriter vw1;

  Size dsize = Size(300, 300);

  Mat img2 = Mat(dsize, CV_32S);

  Mat frame,CloseimgROI,CloseImgLeft;

  chrono::steady_clock::time_point Tbegin, Tend;

  net = dnn::readNetFromTorch("csv_retinanet_5.pt");

  if (net.empty()){

    cout << "init the model net error";

    exit(-1);

  }





  VideoCapture capture;

  frame= capture.open("./2021-01-06-14-22-02-L.mp4");



   if(!capture.isOpened())

  {

    printf("can not open ...\n");

    return -1;

  }

  vw1.open("save.mp4", VideoWriter::fourcc('m', 'p', '4', 'v'), 16, Size(640, 480));

  if (!vw1.isOpened())

    {

       cout << "open vw1 faild" << endl;

        return -1;

    }

  //cout << "Switched to " << (cv::ocl::useOpenCL() ? "OpenCL enabled" : "CPU") << endl;

  //net.setPreferableTarget(DNN_TARGET_OPENCL);

  cout << "Start grabbing, press ESC on Live window to terminate" << endl;

  while(capture.read(frame)){

    //frame=imread("2.jpg");  //need to refresh frame before dnn class detection



    CloseimgROI = changeImgLeft(frame, CloseimgROI);

    CloseImgLeft = CloseimgROI(Range(168, 289), cv::Range(83, 640));//将核心区域裁剪为矩形

    

    

    count++;



    Tbegin = chrono::steady_clock::now();

    detect_from_video(CloseImgLeft,x1,y1,x2,y2,det_index);

    Tend = chrono::steady_clock::now();

    //calculate CloseImgLeft rate

    f = chrono::duration_cast <chrono::milliseconds> (Tend - Tbegin).count();

    if(f>0.0) FPS[((Fcnt++)&0x0F)]=1000.0/f;

    for(f=0.0, i=0;i<16;i++){ f+=FPS[i]; }

    putText(frame, format("FPS %0.2f", f/16),Point(10,20),FONT_HERSHEY_SIMPLEX,0.6, Scalar(0, 0, 255));

    //show output

    imshow("CloseImgLeft", CloseImgLeft);

    

    Rect rec((int)x1+112, (int)y1+205, (int)(x2 - x1), (int)(y2 - y1));

    rectangle(frame,rec, Scalar(0, 0, 255), 2, 8, 0);

    putText(frame, format("%s", class_video_Names[det_index]), Point(x1+112, y1-5+205) ,FONT_HERSHEY_SIMPLEX,1.0, Scalar(0, 0, 255), 2, 8, 0);

     resize(CloseImgLeft, img2, dsize);

     imshow("frame", frame);

     if(count==1550  )

    {

      //vw1 << frame;

       cv::imwrite("save.jpg", frame);

     }

     

    x1=0.0;

    y1=0.0;

    x2=0.0;

    y2=0.0;    

     det_index = 0.0;



    char esc = waitKey(5);

    if(esc == 27) break;

  }

  vw1.release();

  cout << "Closing the camera" << endl;

  destroyAllWindows();

  cout << "Bye!" << endl;

  return 0;

}