retinaface caffe版本後處理的使用教程

環境要求：CUDA9.0（與你的編譯caffe時使用的cuda版本保持一致），opencv3.1（注意博主親試3.4和4.0均有錯，3.4可以編譯但結果不對並不能resize），ubuntu16.04

具體後處理code使用https://github.com/wzj5133329/retinaface_caffe 提供。

1. 首先準備好一個正常的編譯好的可運行的caffe，我這裏參考 https://github.com/eric612/MobileNet-YOLO 提供的版本，您可指定任意可以正常運行的caffe包，推薦https://github.com/weiliu89/caffe 。將caffe編譯好（編譯方法網上教程一籮筐）

2. 安裝opencv3.1.0 source release，具體安裝方法請百度。

3. 下載retinaface_caffe code並解壓，隨意路徑放置，無需放置在caffe目錄下。

4. 按如下參考修改makefile

.PHONY: all test clean deps tags 

CXX=g++
CXXFLAGS += -g -Wall -O -std=c++11                 

原始爲：OPENCVLIBS = `pkg-config opencv --cflags --libs`(即在環境變量pkg-config指定種獲取opencv)
修改爲（具體的指定）：
OPENCVLIBS = -L/home/sh00245/opencv/opencv3.1.0/lib -lopencv_cudabgsegm -lopencv_cudaobjdetect -lopencv_cudastereo -lopencv_shape -lopencv_stitching -lopencv_cudafeatures2d -lopencv_superres -lopencv_cudacodec -lopencv_videostab -lopencv_cudaoptflow -lopencv_cudalegacy -lopencv_calib3d -lopencv_features2d -lopencv_objdetect -lopencv_highgui -lopencv_videoio -lopencv_photo -lopencv_imgcodecs -lopencv_cudawarping -lopencv_cudaimgproc -lopencv_cudafilters -lopencv_video -lopencv_ml -lopencv_imgproc -lopencv_flann -lopencv_cudaarithm -lopencv_core -lopencv_cudev
原始爲：DEPS_INCLUDE_PATH= $(DLIB_PATH)  -I /usr/local/cuda-10.0/include/ -I /home/asd/Project/MobileNet-YOLO-master2/include 
修改爲：（添加opencv頭文件路徑，添加caffe相關頭文件路徑）
DEPS_INCLUDE_PATH= $(DLIB_PATH) -I /home/sh00245/opencv/opencv3.1.0/include/opencv -I /home/sh00245/opencv/opencv3.1.0/include -I /opt/cuda/cuda-9.0_cudnn75/include/ -I /home/sh00245/py2env/MobileNet-YOLO-app/include -I /home/sh00245/py2env/MobileNet-YOLO-app/build/include



TARGET = retinaface


原始爲：LIBS=  -lboost_system -lcaffe -lglog  -lprotobuf -lcudart  -lgflags
修改爲：（添加cuda lib庫和caffe lib庫文件路徑）
LIBS=  -lboost_system -lcaffe -lglog  -lprotobuf -lcudart  -lgflags -L /home/sh00245/py2env/MobileNet-YOLO-app/build/lib -L /opt/cuda/cuda-9.0_cudnn75/lib64

OBJS := $(patsubst %.cpp,%.o,$(wildcard *.cpp))     

  
$(TARGET): $(OBJS) 
	$(CXX)  -o $@  $^ $(LIBS)  $(OPENCVLIBS)  $(DEPS_LIB_PATH) 


%.o:%.cpp
	$(CXX) -c $(CXXFLAGS)  $< $(DEPS_INCLUDE_PATH) 


clean:
	rm -f *.o $(TARGET)

5. 修改main.cpp

#include"detect.h"

using namespace std;

int main(int argc, char** argv) {

    const string proto="mnet.prototxt";
    const string model="mnet.caffemodel";
    const float confidence=0.5;
    const float nms_threshold = 0.4;
    const string cpu_mode="gpu";
    Detector detector(proto,model,confidence,nms_threshold,cpu_mode);
    
    cv::VideoCapture video_frame;
    //video_frame.open(0);
    cv::Mat img;
    //while(video_frame.read(img))
       
        //cv::Mat input = img.clone();
    img=cv::imread("demo.jpg");
    cout<<"channels"<<img.channels()<<endl;
    //cv::resize(img,img,cv::Size(1920,1080));
    cout<<"img10"<<img.rows<<endl;
    uint64_t time1=current_timestamp();
    std::vector<Anchor> result = detector.Detect(img);       
    for(int i = 0; i < result.size(); i ++) {
        cv::rectangle(img, cv::Point((int)result[i].finalbox.x, (int)result[i].finalbox.y), cv::Point((int)result[i].finalbox.width, (int)result[i].finalbox.height), cv::Scalar(0, 255, 255), 2, 8, 0);
        
        cv::circle(img, result[i].pts[0], 10, cv::Scalar(0, 255, 0),-1);
        cv::circle(img, result[i].pts[1], 10, cv::Scalar(0, 255, 0),-1);
        cv::circle(img, result[i].pts[2], 10, cv::Scalar(0, 255, 0),-1);
        cv::circle(img, result[i].pts[3], 10, cv::Scalar(0, 255, 0),-1);
        cv::circle(img, result[i].pts[4], 10, cv::Scalar(0, 255, 0),-1);
    }

    uint64_t time2=current_timestamp();
    cout<<"time:"<<time2-time1<<endl;
    cv::namedWindow("show",CV_WINDOW_NORMAL);
    cv::resizeWindow("show",960,540);
    cv::imshow("show",img);
    cv::waitKey(0);
    cv::imwrite("result.jpg", img);
    return 0;
}

去掉死循環while(true)。

6. cd到retinaface_caffe下，執行make指令

7. 編譯完成後再retinaface_caffe路徑下執行指令./retinaface

以上過程嚴格執行應當可以順利編譯。

記錄之前遇到的錯誤：

a. ‘cv::Rect2f’ has not been declared
error: ‘Rect2f’ in namespace ‘cv’ does not name a type

opencv版本的問題，3.0以下會有這種問題

b. 編譯成功執行時Segmentation fault (core dumped)

opencv使用3.4版本編譯成功，但執行時會報此錯誤，原因是detect.cpp中執行resize後不能正確輸出的問題，若去掉resize，在外部先resize到prototxt中定義的大小後再執行code，能運行並避免該錯誤，但不能得到正確的結果。更換opencv3.1後解決。