face aligned 记录

最新推荐文章于 2024-06-17 16:34:15 发布
最新推荐文章于 2024-06-17 16:34:15 发布
阅读量950 收藏
点赞数
分类专栏： caffe
版权声明：本文为博主原创文章，遵循 CC 4.0 BY-SA 版权协议，转载请附上原文出处链接和本声明。
本文链接：https://blog.csdn.net/u012235274/article/details/53282595
版权

face aligned github
该博客是转化这个里面的matlab程序为c++程序：
转换这个时候感觉遇到一个坑（不知道作者是不是故意这样的），传进去的图片要进行翻转，因为作者在写matlab的时候没有进行翻转就直接放进Blob中，直接导致图片有一个翻转，所以我在c++程序也进行了一个翻转。
现在只完成到第二个网络，后续再更新（已经更新，作者很懒，懒得把其中重复很多的代码写成函数，因此看起来冗余代码很多）。

#include "caffe/net.hpp"
#include <vector>
#include <string>
#include <fstream>
#include <opencv2/opencv.hpp>

void loaddata(boost::shared_ptr<caffe::Net<float> >& net, cv::Mat image);
void loaddata(boost::shared_ptr<caffe::Net<float> >& net, cv::Mat image, int batch_idx);
float maxf(float x1, float x2);
float minf(float x1, float x2);

int main(int argc, char** argv){
    std::string net_path1("/home/yang/MTCNN_face_detection_alignment/code/codes/MTCNNv1/model/det1.prototxt");
    std::string net_path2("/home/yang/MTCNN_face_detection_alignment/code/codes/MTCNNv1/model/det2.prototxt");
    std::string net_path3("/home/yang/MTCNN_face_detection_alignment/code/codes/MTCNNv1/model/det3.prototxt");
    std::string w_path1("/home/yang/MTCNN_face_detection_alignment/code/codes/MTCNNv1/model/det1.caffemodel");
    std::string w_path2("/home/yang/MTCNN_face_detection_alignment/code/codes/MTCNNv1/model/det2.caffemodel");
    std::string w_path3("/home/yang/MTCNN_face_detection_alignment/code/codes/MTCNNv1/model/det3.caffemodel");
    caffe::Caffe::set_mode(caffe::Caffe::GPU);
    boost::shared_ptr<caffe::Net<float> > net1(new caffe::Net<float>(net_path1, caffe::TEST));
    boost::shared_ptr<caffe::Net<float> > net2(new caffe::Net<float>(net_path2, caffe::TEST));
    boost::shared_ptr<caffe::Net<float> > net3(new caffe::Net<float>(net_path3, caffe::TEST));
    net1->CopyTrainedLayersFromBinaryProto(w_path1);
    net2->CopyTrainedLayersFromBinaryProto(w_path2);
    net3->CopyTrainedLayersFromBinaryProto(w_path3);

    //argv[1]表示要进行测试图片的路径txt
    std::ifstream in(argv[1]);
    //std::ifstream in("/home/yang/MTCNN_face_detection_alignment/code/codes/MTCNNv1/model/image.list");
    std::string temp_image_path;
    std::vector<std::string> image_paths;
    //std::cout << "image_path" << std::endl;
    while(in >> temp_image_path){
        image_paths.push_back(temp_image_path);
        //std::cout << temp_image_path << std::endl;
    }
    in.close();
    int image_num = image_paths.size();

    //param of face aligned
    float factor = 0.79;
    int min_size = 10;//min_size of face
    std::vector<float> threshold;
    threshold.push_back(0.9);
    threshold.push_back(0.8);
    threshold.push_back(0.7);
    std::vector<float> scales;

    for(int image_idx = 0; image_idx < image_num; ++image_idx){
        cv::Mat image = cv::imread(image_paths[image_idx]);
        std::cout << "process "<<image_idx <<" " << image_paths[image_idx] << std::endl;
        /*if(image_idx == 0){
            std::cout << "image data " << std::endl;
            for(int tt = 0; tt < 100; ++tt){
                std::cout << int(image.ptr<uchar>(0)[tt*3+2]) << " ";
            }
            std::cout << std::endl;
        }*/
        //std::cout << "image size " << image.rows << " " << image.cols << std::endl;
        scales.clear();
        float m = 12.0/min_size;
        float minl = image.cols < image.rows ? image.cols : image.rows;
        int count=0;
        //scales.push_back(1.0);
        while(minl > 12){
            scales.push_back(m*pow(factor,count));
            minl = minl*factor;
            ++count;
        }
        //std::cout << "scales size = " << scales.size() << std::endl;
        /*for(int i = 0; i < scales.size(); ++i){
            std::cout << scales[i] << " ";
        }*/
        //std::cout << std::endl;
        std::vector<std::vector<float> > total_boxes;
        for(int scales_idx = 0; scales_idx < scales.size(); ++scales_idx){
            int hs = ceil(image.rows * scales[scales_idx]);
            int ws = ceil(image.cols * scales[scales_idx]);
            //std::cout << "resize hs ws " << hs << " "<< ws << std::endl;
            cv::Mat image_s;
            cv::resize(image, image_s, cv::Size(ws, hs));
            //std::cout << "resize image hs ws " << image_s.rows << " "<< image_s.cols << std::endl;
            /*if(scales_idx == 0){
                std::cout << "resize image data " << std::endl;
                for(int tt = 0; tt < 100; ++tt){
                    std::cout << int(image_s.ptr<uchar>(0)[tt*3+2]) << " ";
                }
                std::cout << std::endl;
            }*/
            std::vector<int> shape;
            shape.push_back(1);
            shape.push_back(3);
            shape.push_back(ws);
            shape.push_back(hs);
            net1->input_blobs()[0]->Reshape(shape);
            loaddata(net1, image_s);
            /*if(scales_idx == 0){
                std::ifstream data_in("/home/yang/MTCNN_face_detection_alignment/code/codes/MTCNNv1/test2.txt");
                int tt_temp;
                for(int tt = 0; tt < 431*450*3; ++tt){
                    data_in >> tt_temp;
                    int t1 = tt/(431*450);
                    int t2 = (tt - t1*431*450)/450;
                    int t3 = (tt - t1*431*450)%450;
                    //if(tt < 450)
                    //  std::cout << t1 << " " << t2 << " " << t3 << " " << t1*431*450+t3*450+t2 <<std::endl;
                    net1->input_blobs()[0]->mutable_cpu_data()[t1*431*450+t3*431+t2] = (tt_temp-127.5)*0.0078125;
                }
                data_in.close();
                std::cout << "end data    "<<net1->input_blobs()[0]->mutable_cpu_data()[431*450*3 - 1] <<  " " << net1->input_blobs()[0]->mutable_cpu_data()[431*450*3 - 2] <<" " << net1->input_blobs()[0]->mutable_cpu_data()[431*450*3 - 3] << " " << net1->input_blobs()[0]->mutable_cpu_data()[431*450*3 - 4] << std::endl;
            }
            if(scales_idx == 0){
                std::cout << "image data 111111111 " << std::endl;
                std::cout << (148-127.5)*0.0078125 << std::endl;;
                for(int tt = 0; tt < 100; ++tt)
                    std::cout << net1->input_blobs()[0]->cpu_data()[tt] << " ";
                std::cout << std::endl;
            }*/
            net1->Forward();
            caffe::Blob<float>* coord = net1->output_blobs()[0];
            caffe::Blob<float>* conf = net1->output_blobs()[1];
            /*std::cout << "confidence data " << std::endl;
            for(int tt= 0; tt < 100; ++tt){
                std::cout << conf->cpu_data()[tt] << " ";
            }
            std::cout << std::endl;
            for(;;);
            std::cout << std::endl;*/
            const float* coord_cpu_data = coord->cpu_data();
            const float* conf_cpu_data = conf->cpu_data()+conf->height()*conf->width();
            //std::cout << net1->output_blobs()[0]->num() << " "<< net1->output_blobs()[0]->channels() << " "<< net1->output_blobs()[0]->height() << " "<< net1->output_blobs()[0]->width() << " "<<std::endl;
            //std::cout << net1->output_blobs()[1]->num() << " "<< net1->output_blobs()[1]->channels() << " "<< net1->output_blobs()[1]->height() << " "<< net1->output_blobs()[1]->width() << " "<<std::endl;
            std::vector<std::vector<float> > boxes;
            //std::cout << "output size " << "(" << conf->height() << "," << conf->width() << ")" << std::endl;
            for(int i = 0; i < conf->height(); ++i){
                for(int j = 0; j < conf->width(); ++j){
                    if(conf_cpu_data[i*conf->width()+j] > threshold[0]){
                        //std::cout << "(" << i << "," << j <<")" << " ";
                        std::vector<float> temp;
                        temp.push_back(ceil((2*i)/scales[scales_idx]));//表示宽方向的坐标
                        temp.push_back(ceil((2*j)/scales[scales_idx]));//表示高方向的坐标
                        temp.push_back(ceil((2*i+11)/scales[scales_idx]));
                        temp.push_back(ceil((2*j+11)/scales[scales_idx]));
                        temp.push_back(conf_cpu_data[i*conf->width()+j]);
                        temp.push_back(coord_cpu_data[0*coord->height()*coord->width()+i*coord->width()+j]);
                        temp.push_back(coord_cpu_data[1*coord->height()*coord->width()+i*coord->width()+j]);
                        temp.push_back(coord_cpu_data[2*coord->height()*coord->width()+i*coord->width()+j]);
                        temp.push_back(coord_cpu_data[3*coord->height()*coord->width()+i*coord->width()+j]);

                        boxes.push_back(temp);
                    }
                }
            }
            //std::cout << std::endl;
            for(int i = 0; i < boxes.size(); ++i){
                for(int j = boxes.size()-1; j > i; --j){
                    float temp1 = maxf(boxes[i][0], boxes[j][0]);
                    float temp2 = minf(boxes[i][2], boxes[j][2]);
                    float temp3 = maxf(boxes[i][1], boxes[j][1]);
                    float temp4 = minf(boxes[i][3], boxes[j][3]);
                    temp1 = temp2 - temp1;
                    temp3 = temp4 - temp3;
                    float iou;
                    if(temp1 < 0 || temp3 < 0){
                        iou = 0.0;
                    }else{
                        float area1 = (boxes[i][2] - boxes[i][0])*(boxes[i][3] - boxes[i][1]);
                        float area2 = (boxes[j][2] - boxes[j][0])*(boxes[j][3] - boxes[j][1]);
                        iou = temp1*temp3/(area1 + area2 - temp1*temp3);
                    }
                    if(iou > 0.5){
                        //std::cout << boxes[i][4]  << " " << boxes[j][4] << std::endl;
                        if(boxes[i][4] > boxes[j][4]){
                            boxes.erase(boxes.begin()+j);
                        }else{
                            boxes.erase(boxes.begin()+i);
                            --i;
                            break;
                        }
                    }
                }
            }
            //std::cout << boxes.size() << std::endl;
            for(int k = 0; k < boxes.size(); ++k){
                total_boxes.push_back(boxes[k]);
            }
        }
        for(int i = 0; i < total_boxes.size(); ++i){
            for(int j = total_boxes.size()-1; j > i; --j){
                float temp1 = maxf(total_boxes[i][0], total_boxes[j][0]);
                float temp2 = minf(total_boxes[i][2], total_boxes[j][2]);
                float temp3 = maxf(total_boxes[i][1], total_boxes[j][1]);
                float temp4 = minf(total_boxes[i][3], total_boxes[j][3]);
                temp1 = temp2 - temp1;
                temp3 = temp4 - temp3;
                float iou;
                if(temp1 < 0 || temp3 < 0){
                    iou = 0.0;
                }else{
                    float area1 = (total_boxes[i][2] - total_boxes[i][0])*(total_boxes[i][3] - total_boxes[i][1]);
                    float area2 = (total_boxes[j][2] - total_boxes[j][0])*(total_boxes[j][3] - total_boxes[j][1]);
                    iou = temp1*temp3/(area1 + area2 - temp1*temp3);
                }
                if(iou > 0.7){
                    if(total_boxes[i][4] > total_boxes[j][4]){
                        total_boxes.erase(total_boxes.begin()+j);
                    }else{
                        total_boxes.erase(total_boxes.begin()+i);
                        --i;
                        break;
                    }
                }
            }
        }
        for(int i = 0; i < total_boxes.size(); ++i){
            float regh = total_boxes[i][2] - total_boxes[i][0];
            float regw = total_boxes[i][3] - total_boxes[i][1];
            total_boxes[i][0] += total_boxes[i][5]*regh;
            if(total_boxes[i][0] < 0) total_boxes[i][0] = 0;
            total_boxes[i][1] += total_boxes[i][6]*regw;
            if(total_boxes[i][1] < 0) total_boxes[i][1] = 0;
            total_boxes[i][2] += total_boxes[i][7]*regh;
            //if(total_boxes[i][2] > image.cols) total_boxes[i][2] = image.cols;
            total_boxes[i][3] += total_boxes[i][8]*regw;
            //if(total_boxes[i][3] > image.rows) total_boxes[i][3] = image.rows;
            int max_coord = maxf(regh, regw);
            total_boxes[i][0] += (regh*0.5-max_coord*0.5);
            total_boxes[i][1] += (regw*0.5-max_coord*0.5);
            total_boxes[i][2] = total_boxes[i][0]+max_coord;
            total_boxes[i][3] = total_boxes[i][1]+max_coord;
            if(total_boxes[i][2] > image.cols) total_boxes[i][2] = image.cols;
            if(total_boxes[i][3] > image.rows) total_boxes[i][3] = image.rows;
            cv::Point point1(total_boxes[i][0], total_boxes[i][1]);
            cv::Point point2(total_boxes[i][2],total_boxes[i][3]);
            //cv::rectangle(image, cv::Rect(point2,point1),cv::Scalar(0,255,255),1.0);
        }
        //cv::imshow("image", image);
        //cv::waitKey(0);
        //std::cout <<"net 1 boxes num = " <<total_boxes.size() << std::endl;
        if(total_boxes.size() == 0) continue;
        std::vector<int> shape;
        shape.push_back(total_boxes.size());
        shape.push_back(3);
        shape.push_back(24);
        shape.push_back(24);
        net2->input_blobs()[0]->Reshape(shape);
        for(int i = 0; i < total_boxes.size(); ++i){
            //std::cout <<total_boxes[i][0] << " " << total_boxes[i][1]  << " " << total_boxes[i][2] << " " << total_boxes[i][3] <<std::endl;
            cv::Point p1(total_boxes[i][0], total_boxes[i][1]);
            cv::Point p2(total_boxes[i][2], total_boxes[i][3]);
            //cv::Mat image_roi = image(cv::Rect(int(total_boxes[i][1]), int(total_boxes[i][0]), int(total_boxes[i][3])-int(total_boxes[i][1]), int(total_boxes[i][2])-int(total_boxes[i][0])));
            cv::Mat image_roi = image(cv::Rect(p1, p2));
            cv::Mat temp;
            //cv::imshow("image_temp", image_roi);
            //cv::waitKey(0);
            cv::resize(image_roi, temp, cv::Size(24,24));
            //cv::imshow("image_temp", temp);
            //cv::waitKey(0);
            //std::cout << temp.rows << " " << temp.cols << std::endl;
            loaddata(net2, temp, i);
        }
        net2->Forward();
        //std::cout << net2->output_blobs()[0]->num() << " "<< net2->output_blobs()[0]->channels() << " "<< net2->output_blobs()[0]->height() << " "<< net2->output_blobs()[0]->width() << " "<<std::endl;
        //std::cout << net2->output_blobs()[1]->num() << " "<< net2->output_blobs()[1]->channels() << " "<< net2->output_blobs()[1]->height() << " "<< net2->output_blobs()[1]->width() << " "<<std::endl;
        const float* conf_data = net2->output_blobs()[1]->cpu_data();
        const float* coor_data = net2->output_blobs()[0]->cpu_data();
        for(int i = net2->output_blobs()[0]->num()-1 ; i >= 0; --i){
            if(conf_data[i*2 + 1] >= threshold[1]){
                total_boxes[i][4] = conf_data[i*2+1];
                total_boxes[i][5] = coor_data[i*4+0];
                total_boxes[i][6] = coor_data[i*4+1];
                total_boxes[i][7] = coor_data[i*4+2];
                total_boxes[i][8] = coor_data[i*4+3];
            }else{
                total_boxes.erase(total_boxes.begin()+i);
            }
        }
        for(int i = 0; i < total_boxes.size(); ++i){
            float regh = total_boxes[i][2] - total_boxes[i][0];
            float regw = total_boxes[i][3] - total_boxes[i][1];
            total_boxes[i][0] += total_boxes[i][5]*regh;
            if(total_boxes[i][0] < 0) total_boxes[i][0] = 0;
            total_boxes[i][1] += total_boxes[i][6]*regw;
            if(total_boxes[i][1] < 0) total_boxes[i][1] = 0;
            total_boxes[i][2] += total_boxes[i][7]*regh;
            total_boxes[i][3] += total_boxes[i][8]*regw;


            /*int max_coord = maxf(regh, regw);
            total_boxes[i][0] += (regh*0.5-max_coord*0.5);
            total_boxes[i][1] += (regw*0.5-max_coord*0.5);
            total_boxes[i][2] = total_boxes[i][0]+max_coord;
            total_boxes[i][3] = total_boxes[i][1]+max_coord;*/

            if(total_boxes[i][2] > image.cols) total_boxes[i][2] = image.cols;
            if(total_boxes[i][3] > image.rows) total_boxes[i][3] = image.rows;

        }
        for(int i = 0; i < total_boxes.size(); ++i){
            for(int j = total_boxes.size()-1; j > i; --j){
                float temp1 = maxf(total_boxes[i][0], total_boxes[j][0]);
                float temp2 = minf(total_boxes[i][2], total_boxes[j][2]);
                float temp3 = maxf(total_boxes[i][1], total_boxes[j][1]);
                float temp4 = minf(total_boxes[i][3], total_boxes[j][3]);
                temp1 = temp2 - temp1;
                temp3 = temp4 - temp3;
                float iou;
                if(temp1 < 0 || temp3 < 0){
                    iou = 0.0;
                }else{
                    float area1 = (total_boxes[i][2] - total_boxes[i][0])*(total_boxes[i][3] - total_boxes[i][1]);
                    float area2 = (total_boxes[j][2] - total_boxes[j][0])*(total_boxes[j][3] - total_boxes[j][1]);
                    //float area3 = minf(area1, area2);
                    iou = temp1*temp3/(area1+area2-temp1*temp3);
                }
                if(iou > 0.7){
                    if(total_boxes[i][4] > total_boxes[j][4]){
                        total_boxes.erase(total_boxes.begin()+j);
                    }else{
                        total_boxes.erase(total_boxes.begin()+i);
                        --i;
                        break;
                    }
                }
            }
        }
        for(int i = 0; i < total_boxes.size(); ++i){

            cv::Point point1(total_boxes[i][0], total_boxes[i][1]);
            cv::Point point2(total_boxes[i][2],total_boxes[i][3]);

            //cv::rectangle(image, cv::Rect(point2,point1),cv::Scalar(0,255,255),1.0);
        }
        //cv::imshow("image", image);
        //cv::waitKey(0);
        //std::cout << "net 2 process boxes num = " << total_boxes.size() << std::endl;
        if(total_boxes.size() == 0) continue;
        std::vector<int> shape2;
        shape2.push_back(total_boxes.size());
        shape2.push_back(3);
        shape2.push_back(48);
        shape2.push_back(48);
        net3->input_blobs()[0]->Reshape(shape2);
        for(int i = 0; i < total_boxes.size(); ++i){
            cv::Point point1(total_boxes[i][0], total_boxes[i][1]);
            cv::Point point2(total_boxes[i][2],total_boxes[i][3]);
            cv::Mat face_need_align = image(cv::Rect(point1, point2));
            cv::resize(face_need_align, face_need_align, cv::Size(48,48));
            loaddata(net3, face_need_align, i);
        }
        net3->Forward();
        for(int i = total_boxes.size()-1; i >= 0; --i){
            if(net3->output_blobs()[2]->cpu_data()[i*2+1] < threshold[2]){
                total_boxes.erase(total_boxes.begin()+i);
            }else{
                float w = total_boxes[i][2] - total_boxes[i][0];
                float h = total_boxes[i][3] - total_boxes[i][1];
                total_boxes[i][5] = net3->output_blobs()[0]->cpu_data()[i*4+0];
                total_boxes[i][6] = net3->output_blobs()[0]->cpu_data()[i*4+1];
                total_boxes[i][7] = net3->output_blobs()[0]->cpu_data()[i*4+2];
                total_boxes[i][8] = net3->output_blobs()[0]->cpu_data()[i*4+3];
                total_boxes[i].push_back(net3->output_blobs()[1]->cpu_data()[i*10+0]*w+total_boxes[i][0]);
                total_boxes[i].push_back(net3->output_blobs()[1]->cpu_data()[i*10+1]*w+total_boxes[i][0]);
                total_boxes[i].push_back(net3->output_blobs()[1]->cpu_data()[i*10+2]*w+total_boxes[i][0]);
                total_boxes[i].push_back(net3->output_blobs()[1]->cpu_data()[i*10+3]*w+total_boxes[i][0]);
                total_boxes[i].push_back(net3->output_blobs()[1]->cpu_data()[i*10+4]*w+total_boxes[i][0]);
                total_boxes[i].push_back(net3->output_blobs()[1]->cpu_data()[i*10+5]*h+total_boxes[i][1]);
                total_boxes[i].push_back(net3->output_blobs()[1]->cpu_data()[i*10+6]*h+total_boxes[i][1]);
                total_boxes[i].push_back(net3->output_blobs()[1]->cpu_data()[i*10+7]*h+total_boxes[i][1]);
                total_boxes[i].push_back(net3->output_blobs()[1]->cpu_data()[i*10+8]*h+total_boxes[i][1]);
                total_boxes[i].push_back(net3->output_blobs()[1]->cpu_data()[i*10+9]*h+total_boxes[i][1]);
            }
        }

        for(int i = 0; i < total_boxes.size(); ++i){
            for(int j = total_boxes.size()-1; j > i; --j){
                float temp1 = maxf(total_boxes[i][0], total_boxes[j][0]);
                float temp2 = minf(total_boxes[i][2], total_boxes[j][2]);
                float temp3 = maxf(total_boxes[i][1], total_boxes[j][1]);
                float temp4 = minf(total_boxes[i][3], total_boxes[j][3]);
                temp1 = temp2 - temp1;
                temp3 = temp4 - temp3;
                float iou;
                if(temp1 < 0 || temp3 < 0){
                    iou = 0.0;
                }else{
                    float area1 = (total_boxes[i][2] - total_boxes[i][0])*(total_boxes[i][3] - total_boxes[i][1]);
                    float area2 = (total_boxes[j][2] - total_boxes[j][0])*(total_boxes[j][3] - total_boxes[j][1]);
                    float area3 = minf(area1, area2);
                    iou = temp1*temp3/(area3);
                }
                if(iou > 0.7){
                    if(total_boxes[i][4] > total_boxes[j][4]){
                        total_boxes.erase(total_boxes.begin()+j);
                    }else{
                        total_boxes.erase(total_boxes.begin()+i);
                        --i;
                        break;
                    }
                }
            }
        }

        for(int i = 0; i < total_boxes.size(); ++i){
            //还没进行回归检测

            float regh = total_boxes[i][2] - total_boxes[i][0];
            float regw = total_boxes[i][3] - total_boxes[i][1];
            total_boxes[i][0] += total_boxes[i][5]*regh;
            if(total_boxes[i][0] < 0) total_boxes[i][0] = 0;
            total_boxes[i][1] += total_boxes[i][6]*regw;
            if(total_boxes[i][1] < 0) total_boxes[i][1] = 0;
            total_boxes[i][2] += total_boxes[i][7]*regh;
            total_boxes[i][3] += total_boxes[i][8]*regw;


            /*int max_coord = maxf(regh, regw);
            total_boxes[i][0] += (regh*0.5-max_coord*0.5);
            total_boxes[i][1] += (regw*0.5-max_coord*0.5);
            total_boxes[i][2] = total_boxes[i][0]+max_coord;
            total_boxes[i][3] = total_boxes[i][1]+max_coord;*/

            cv::Point point1(total_boxes[i][0], total_boxes[i][1]);
            cv::Point point2(total_boxes[i][2],total_boxes[i][3]);

            cv::rectangle(image, cv::Rect(point2,point1),cv::Scalar(0,255,255),1.0);
            cv::rectangle(image, cv::Rect(total_boxes[i][9], total_boxes[i][14], 1, 1), cv::Scalar(255,255,0), 2.0);
            cv::rectangle(image, cv::Rect(total_boxes[i][10], total_boxes[i][15], 1, 1), cv::Scalar(255,255,0), 2.0);
            cv::rectangle(image, cv::Rect(total_boxes[i][11], total_boxes[i][16], 1, 1), cv::Scalar(255,255,0), 2.0);
            cv::rectangle(image, cv::Rect(total_boxes[i][12], total_boxes[i][17], 1, 1), cv::Scalar(255,255,0), 2.0);
            cv::rectangle(image, cv::Rect(total_boxes[i][13], total_boxes[i][18], 1, 1), cv::Scalar(255,255,0), 2.0);
        }
        cv::imshow("image", image);
        cv::waitKey(1);
    }
    return 1;
}
void loaddata(boost::shared_ptr<caffe::Net<float> >& net, cv::Mat image){
  caffe::Blob<float>* input_layer = net->input_blobs()[0];
  int width, height;
  width = input_layer->width();
  height = input_layer->height();
  int size = width*height;
  float* input_data = input_layer->mutable_cpu_data();
  int temp,idx;
  for(int i = 0; i < width; ++i){
    uchar* pdata = image.ptr<uchar>(i);
    for(int j = 0; j < height; ++j){
      temp = 3*j;
      idx = j*width+i;
      input_data[idx] = (pdata[temp+2]-127.5)*0.0078125;
      input_data[idx+size] = (pdata[temp+1]-127.5)*0.0078125;
      input_data[idx+2*size] = (pdata[temp+0]-127.5)*0.0078125;
    }
  }
}
void loaddata(boost::shared_ptr<caffe::Net<float> >& net, cv::Mat image, int batch_idx){
  caffe::Blob<float>* input_layer = net->input_blobs()[0];
  int width, height;
  width = input_layer->width();
  height = input_layer->height();
  int size = width*height;
  float* input_data = input_layer->mutable_cpu_data() + batch_idx*size*3;
  int temp,idx;
  for(int i = 0; i < width; ++i){
    uchar* pdata = image.ptr<uchar>(i);
    for(int j = 0; j < height; ++j){
      temp = 3*j;
      idx = j*width+i;
      input_data[idx] = (pdata[temp+2]-127.5)*0.0078125;
      input_data[idx+size] = (pdata[temp+1]-127.5)*0.0078125;
      input_data[idx+2*size] = (pdata[temp+0]-127.5)*0.0078125;
    }
  }
}
float maxf(float x1, float x2){
    if(x1 > x2)
        return x1;
    else
        return x2;
}
float minf(float x1, float x2){
    if(x1 > x2)
        return x2;
    else
        return x1;
}
小_小_杨_
关注
专栏目录
Insightface/Arcface项目实践流程
Zebiao Wu Blog
08-20 1万+
Insightface/Arcface项目实践流程 Code:https://github.com/deepinsight/insightface 作者提供了完整的工程，能够基本满足并完成人脸识别的这个流程。 以下纯属个人实践记录，如有不当还望各位前辈多多指教。 人脸识别流程4步：1、检测；2、对齐矫正；3、提取特征；特征匹配 其中，检测对齐使用ssh或mtcnn并用dlib即可实现...
作业记录 WEEK6
Sycret
03-06 348
消融实验：对比不同embedding dim时模型准确率 在上次的作业基础上，更换训练数据集，添加测试数据集与代码。 步骤 0. 环境 pip install cmake pip install dlib 要等很久，尝试了永久安装路径，cmake可以装到永久路径，但是dlib一装内存就溢出，环境就被关闭了。也不知道为什么。 ！换了一下环境到GPU，发现好像可以了诶，问题解决。 1. 训练集 训练数据集选择：VGGFACE2 的训练集 或者 VGGFACE2的测试集 - 下载完成vggface2_tra
MTCNN_face_detection_alignment.zip
10-13
ＭＴＣＮＮ的ｍａｔｌａｂ程序，稍加改动，据说比作者的好点
Aligning face images
todayq的专栏
11-19 2265
Aligning face images Some people have asked me how I've aligned the face images in my articles. Giving people my ImageMagick hack is embarrassing for me, so I've decided to rewrite it into a Python s
人脸识别系统---年龄预测
最新发布
2303_77278243的博客
06-17 670
2.加载预训练的性别和年龄识别模型 3.定义性别和年龄的标签列表 4.定义选择图片的函数 5.创建一个按钮，用于打开文件选择对话框 6.定义显示图片的函数 6.1确保 img 变量已经定义 6.2调整图像大小到相等的大小 6.3将 OpenCV 图像转换为 PIL 图像 6.4在这里调整图像到固定大小 6.5将 PIL 图像转换为 tkinter 支持的格式 6.6在 root 窗口中创建一个标签来显示图像 7.创建预测年龄的函数 7.1确保 img 变量已经定义 7.2转换为灰度
人脸识别 (4) 人脸对齐
light169的专栏
04-22 1607
参考：FaceDetector/face_align.ipynb at master · faciallab/FaceDetector · GitHub 中文：从零开始搭建人脸识别系统（二）：人脸对齐 - 知乎 Face Alignment Step-by-Step 1、Align Faces by Spatial Transform Operation import cv2 import matplotlib.pyplot as plt import numpy as np img_file
计算机视觉知识点-人脸对齐
蓝色的杯子
07-20 804
今天这篇博客介绍了一下人脸对齐的原理，并通过代码演示了一下人脸对齐。我参考了这篇博客
MTCNN解读：Joint Face Detection and Alignment using Multi-task Cascaded Convolutional Networks
热门推荐
VivienFu的专栏
06-14 2万+
解读论文为《Joint Face Detection and Alignment using Multi-task Cascaded Convolutional Networks》 基于多任务级联卷积神经网络的人脸检测和对齐 论文地址： https://kpzhang93.github.io/MTCNN_face_detection_alignment/ 本文作者：
论文笔记1.2——Joint Face Detection and Alignment using Multi-task Cascaded Convolutional Networks
大西瓜不甜的博客
02-28 414
二. APPROACH In this section, we will describe our approach towards joint face detection and alignment. A. Overall Framework The overall pipeline of our approach is shown in Fig. 1. Given an...
A Discriminative Feature Learning Approach for Deep Face Recognition 原理及在caffe实验复现
dongfang1984的博客
11-25 1万+
本文主要讲centerloss 的原理，及其创新点。然后用caffe 跑自己的数据（CASIA-WebFace | MsCelebV1-Faces-Aligned） Reference paper：A Discriminative Feature Learning Approach for Deep Face Recognition ECCV:2016 github:https
RetinaFace：一种野外多层次人脸定位方法
1RetinaFace：野外单次多层次人脸定位邓建康*1，2，3郭佳*2Evangelos Ververas1，3Irene Kotsia4Stefanos Zafeiriou1，31帝国理工学院2InsightFace3 FaceSoft4伦敦米德尔塞克斯大学{j.deng16，e.ververas16，s....
基于OpenCV的人脸对齐步骤详解及源码实现
阿旭的博客
06-25 1万+
基于OpenCV的人脸对齐步骤详解及源码实现
MTCNN（Multi-task convolutional neural networks）人脸对齐
萧海的博客
10-18 656
该MTCNN算法出自深圳先进技术研究院，乔宇老师组，是今年2016的ECCV。（至少我知道的今年已经一篇cvpr，一篇eccv了）。     进入正题 理论基础： 正如上图所示，该MTCNN由3个网络结构组成（P-Net,R-Net,O-Net）。 Proposal Network (P-Net)：该网络结构主要获得了人脸区域的候选窗口和边界框的回归向量。并用该边界框做回归...
Face Detection and Alignment with MTCNN 人脸检测对齐 快速实现 Python代码
JerryZhang__的博客
09-09 3146
论文Joint Face Detection and Alignment using Multi-task Cascaded Convolutional Networks提出了一种多任务-级联的卷积神经网络框架，用以实现人脸检测和对齐。构成人脸检测和对齐的级联网络的三个部分分别为：P-Net，R-Net和O-Net. (1)P-Net：Proposal Net，用于实现人脸候选框提取； (2)...
MTCNN 人脸检测 人脸对齐
flyfish
11-06 1206
MTCNN 人脸检测 人脸对齐 flyfish 总体 Loss Stage 0:Image Pyramid Stage 1:P-Net P-Net流程 输入图像尺寸，定义为(h, w) 最小人脸尺寸，定义为 min_face_size Image Pyramid缩放因子，定义为factor 初始缩放因子定义为init_factor=12/min_face_size P-Net网络的初始大小1...
人脸检测与对齐之MTCNN网络
知识搬运工的博客
06-20 2912
本文将对人脸检测MTCNN算法做一个简要介绍。该MTCNN算法出自深圳先进技术研究院乔宇老师组，是今年2016的ECCV。 论文原文：https://kpzhang93.github.io/MTCNN_face_detection_alignment/paper/spl.pdf 论文译文：https://blog.csdn.net/lff1208/article/details/7732...
C++版 Face Alignment at 3000FPS（一）TrainModel运行
rongrongyaofeiqi的博客
11-03 3471
C++版 Face Alignment at 3000FPS TrainModel运行
Face - align_dlib.py
It is me 的博客
07-02 261
#Code from：cmusatyalab/openface # Copyright 2015-2016 Carnegie Mellon University # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the Lice.
人脸矫正工程实现
On_theway10的博客
12-06 769
Face Landmark [1]. mtcnn-pytorch|Joint Face Detection and Alignment usingMulti-task Cascaded Convolutional Networks[2016] [2]. mtcnn-blog:https://blog.csdn.net/On_theway10/article/details/1026...
Aligned Reads
06-04
Aligned Reads（比对读段）是指测序数据中已经通过比对算法与参考基因组序列进行了比对，并成功定位到基因组上的测序片段。在基因组测序中，测序仪会生成大量的短读段（通常是50-500bp），这些读段需要通过比对算法...
评论 2
添加红包

请填写红包祝福语或标题

个

红包个数最小为10个

元

红包金额最低5元

当前余额3.43元 前往充值 >
需支付：10.00元
成就一亿技术人!
领取后你会自动成为博主和红包主的粉丝 规则
hope_wisdom
发出的红包
实付元
使用余额支付
点击重新获取
扫码支付
钱包余额 0

抵扣说明：

1.余额是钱包充值的虚拟货币，按照1:1的比例进行支付金额的抵扣。
2.余额无法直接购买下载，可以购买VIP、付费专栏及课程。

余额充值