人脸检测器模型训练方法

01 dlib人脸检测器原始数据获取

68个关键点的训练数据集(1.7GB)：http://dlib.net/files/data/ibug_300W_large_face_landmark_dataset.tar.gz
194个关键点的数据集(需要*)：http://*.com/questions/36711905/dlib-train-shape-predictor-ex-cpp?answertab=votes#tab-top

02 训练代码

dlib中examples中的代码。
dlib/examples/train_shape_predictor_ex.cpp

// dlib/examples/train_shape_predictor_ex.cpp
#include <dlib/image_processing.h>
#include <dlib/data_io.h>
#include <iostream>
using namespace dlib;
using namespace std;
std::vector<std::vector<double> > get_interocular_distances (
const std::vector<std::vector<full_object_detection> >& objects
);
int main(int argc, char** argv)
{
try
{
if (argc != 2)
{
cout << "Give the path to the examples/faces directory as the argument to this" << endl;
cout << "program. For example, if you are in the examples folder then execute " << endl;
cout << "this program by running: " << endl;
cout << " ./train_shape_predictor_ex faces" << endl;
cout << endl;
return 0;
}
const std::string faces_directory = argv[1];
dlib::array<array2d<unsigned char> > images_train, images_test;
std::vector<std::vector<full_object_detection> > faces_train, faces_test;
// 1. 载入训练集，测试集
// load_image_dataset(images_train, faces_train, faces_directory+"/training_with_face_landmarks.xml");
// load_image_dataset(images_test, faces_test, faces_directory+"/testing_with_face_landmarks.xml");
// 68个点的训练数据集:http://dlib.net/files/data/ibug_300W_large_face_landmark_dataset.tar.gz
load_image_dataset(images_train, faces_train, faces_directory + "/labels_ibug_300W_train.xml");
load_image_dataset(images_test, faces_test, faces_directory + "/labels_ibug_300W_test.xml");
shape_predictor_trainer trainer;
// 测试中调节了 tree_depth参数：2,4, 5, 10
// 测试机器为8核，set_num_threads使用8，训练时cpu:70%
trainer.set_oversampling_amount(300);
trainer.set_nu(0.05);
trainer.set_tree_depth(5);
trainer.set_num_threads(8);
trainer.be_verbose();
// 训练
shape_predictor sp = trainer.train(images_train, faces_train);
cout << "mean training error: "<<
test_shape_predictor(sp, images_train, faces_train, get_interocular_distances(faces_train)) << endl;
cout << "mean testing error: "<<
test_shape_predictor(sp, images_test, faces_test, get_interocular_distances(faces_test)) << endl;
// 保存模型
serialize("sp.dat") << sp;
std::string str;
std::cin >> str;
}
catch (exception& e)
{
cout << "\nexception thrown!" << endl;
cout << e.what() << endl;
}
}
double interocular_distance (
const full_object_detection& det
)
{
dlib::vector<double,2> l, r;
double cnt = 0;
for (unsigned long i = 36; i <= 41; ++i)
{
l += det.part(i);
++cnt;
}
l /= cnt;
cnt = 0;
for (unsigned long i = 42; i <= 47; ++i)
{
r += det.part(i);
++cnt;
}
r /= cnt;
return length(l-r);
}
std::vector<std::vector<double> > get_interocular_distances (
const std::vector<std::vector<full_object_detection> >& objects
)
{
std::vector<std::vector<double> > temp(objects.size());
for (unsigned long i = 0; i < objects.size(); ++i)
{
for (unsigned long j = 0; j < objects[i].size(); ++j)
{
temp[i].push_back(interocular_distance(objects[i][j]));
}
}
return temp;
}

03 训练结果

194关键点训练情况：
tree_depth=2，num_threads=2，
Release版本训练时间 5+小时
Debug版本训练时间 148+小时(训练一定要使用Release版本)
tree_depth=2,sp.dat=44.6MB，占用内存最大11GB
tree_depth=10,sp.data=11GB

68关键点训练结果：
tree_depth=2，num_threads=8，CPU:70% 内存：20+GB 6+小时 sp.data=15.8MB
tree_depth=4，num_threads=8，CPU:70% 内存：20+GB 12+小时 sp.data=63.3MB
tree_depth=5，num_threads=8，CPU:70% 内存：20+GB 16+小时 sp.data=126MB

tree_depth=5，num_threads=8
mean training error: 0.0479476
mean testing error: 0.0586204

04 测试效果

使用dlib中examples中的代码测试。
dlib/examples/face_landmark_detection_ex.cpp
也可食用dlib提供的训练模型：
http://dlib.net/files/shape_predictor_68_face_landmarks.dat.bz2

原图：
人脸检测器模型训练方法
检测效果图：

05 参考

http://blog.csdn.net/jcx1314/article/details/65937839
http://blog.csdn.net/elaine_bao/article/details/53054533