人脸检测器模型训练方法（从数据准备到模型部署）

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(需要翻墙)：http://stackoverflow.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

1. // dlib/examples/train_shape_predictor_ex.cpp
2. #include <dlib/image_processing.h>
3. #include <dlib/data_io.h>
4. #include <iostream>
6. using namespace dlib;
7. using namespace std;
9. std::vector<std::vector<double> > get_interocular_distances (
10. const std::vector<std::vector<full_object_detection> >& objects
11. );
13. int main(int argc, char** argv)
14. {
15. try
16. {
17. if (argc != 2)
18. {
19. cout << "Give the path to the examples/faces directory as the argument to this" << endl;
20. cout << "program. For example, if you are in the examples folder then execute " << endl;
21. cout << "this program by running: " << endl;
22. cout << " ./train_shape_predictor_ex faces" << endl;
23. cout << endl;
24. return 0;
25. }
26. const std::string faces_directory = argv[1];
27. dlib::array<array2d<unsigned char> > images_train, images_test;
28. std::vector<std::vector<full_object_detection> > faces_train, faces_test;
29. // 1. 载入训练集，测试集
30. // load_image_dataset(images_train, faces_train, faces_directory+"/training_with_face_landmarks.xml");
31. // load_image_dataset(images_test, faces_test, faces_directory+"/testing_with_face_landmarks.xml");
32. // 68个点的训练数据集:http://dlib.net/files/data/ibug_300W_large_face_landmark_dataset.tar.gz
33. load_image_dataset(images_train, faces_train, faces_directory + "/labels_ibug_300W_train.xml");
34. load_image_dataset(images_test, faces_test, faces_directory + "/labels_ibug_300W_test.xml");
36. shape_predictor_trainer trainer;
38. // 测试中调节了 tree_depth参数：2,4, 5, 10
39. // 测试机器为8核，set_num_threads使用8，训练时cpu:70%
40. trainer.set_oversampling_amount(300);
41. trainer.set_nu(0.05);
42. trainer.set_tree_depth(5);
43. trainer.set_num_threads(8);
44. trainer.be_verbose();
45. // 训练
46. shape_predictor sp = trainer.train(images_train, faces_train);
48. cout << "mean training error: "<<
49. test_shape_predictor(sp, images_train, faces_train, get_interocular_distances(faces_train)) << endl;
51. cout << "mean testing error: "<<
52. test_shape_predictor(sp, images_test, faces_test, get_interocular_distances(faces_test)) << endl;
53. // 保存模型
54. serialize("sp.dat") << sp;
55. std::string str;
56. std::cin >> str;
57. }
58. catch (exception& e)
59. {
60. cout << "\nexception thrown!" << endl;
61. cout << e.what() << endl;
62. }
63. }
65. double interocular_distance (
66. const full_object_detection& det
67. )
68. {
69. dlib::vector<double,2> l, r;
70. double cnt = 0;
72. for (unsigned long i = 36; i <= 41; ++i)
73. {
74. l += det.part(i);
75. ++cnt;
76. }
77. l /= cnt;
79. cnt = 0;
80. for (unsigned long i = 42; i <= 47; ++i)
81. {
82. r += det.part(i);
83. ++cnt;
84. }
85. r /= cnt;
87. return length(l-r);
88. }
90. std::vector<std::vector<double> > get_interocular_distances (
91. const std::vector<std::vector<full_object_detection> >& objects
92. )
93. {
94. std::vector<std::vector<double> > temp(objects.size());
95. for (unsigned long i = 0; i < objects.size(); ++i)
96. {
97. for (unsigned long j = 0; j < objects[i].size(); ++j)
98. {
99. temp[i].push_back(interocular_distance(objects[i][j]));
100. }
101. }
102. return temp;
103. }

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

1. tree_depth=5，num_threads=8
2. mean training error: 0.0479476
3. 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：


   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

免责声明：本文系网络转载或改编，未找到原创作者，版权归原作者所有。如涉及版权，请联系删