| 100 | // ---------------------------------------------------------------------------------------- |
| 101 | |
| 102 | int main() |
| 103 | { |
| 104 | try |
| 105 | { |
| 106 | // The first thing we do is create the set of 3 images discussed above. |
| 107 | dlib::array<array2d<unsigned char> > images; |
| 108 | std::vector<std::vector<rectangle> > object_locations; |
| 109 | make_simple_test_data(images, object_locations); |
| 110 | |
| 111 | |
| 112 | /* |
| 113 | This next block of code specifies the type of sliding window classifier we will |
| 114 | be using to detect the white squares. The most important thing here is the |
| 115 | scan_image_pyramid template. Instances of this template represent the core |
| 116 | of a sliding window classifier. To go into more detail, the sliding window |
| 117 | classifiers used by this object have three parts: |
| 118 | 1. The underlying feature extraction. See the dlib documentation for a detailed |
| 119 | discussion of how the hashed_feature_image and hog_image feature extractors |
| 120 | work. However, to understand this example, all you need to know is that the |
| 121 | feature extractor associates a vector with each location in an image. This |
| 122 | vector is supposed to capture information which describes how parts of the |
| 123 | image look. Importantly, it should do this in a way that is relevant to the |
| 124 | problem you are trying to solve. |
| 125 | |
| 126 | 2. A detection template. This is a rectangle which defines the shape of a |
| 127 | sliding window (i.e. the object_box), as well as a set of rectangular feature |
| 128 | extraction regions inside it. This set of regions defines the spatial |
| 129 | structure of the overall feature extraction within a sliding window. In |
| 130 | particular, each location of a sliding window has a feature vector |
| 131 | associated with it. This feature vector is defined as follows: |
| 132 | - Let N denote the number of feature extraction zones. |
| 133 | - Let M denote the dimensionality of the vectors output by Feature_extractor_type |
| 134 | objects. |
| 135 | - Let F(i) == the M dimensional vector which is the sum of all vectors |
| 136 | given by our Feature_extractor_type object inside the ith feature extraction |
| 137 | zone. |
| 138 | - Then the feature vector for a sliding window is an M*N dimensional vector |
| 139 | [F(1) F(2) F(3) ... F(N)] (i.e. it is a concatenation of the N vectors). |
| 140 | This feature vector can be thought of as a collection of N "bags of features", |
| 141 | each bag coming from a spatial location determined by one of the rectangular |
| 142 | feature extraction zones. |
| 143 | |
| 144 | 3. A weight vector and a threshold value. The dot product between the weight |
| 145 | vector and the feature vector for a sliding window location gives the score |
| 146 | of the window. If this score is greater than the threshold value then the |
| 147 | window location is output as a detection. You don't need to determine these |
| 148 | parameters yourself. They are automatically populated by the |
| 149 | structural_object_detection_trainer. |
| 150 | |
| 151 | The sliding window classifiers described above are applied to every level of an |
| 152 | image pyramid. So you need to tell scan_image_pyramid what kind of pyramid you want |
| 153 | to use. In this case we are using pyramid_down<2> which downsamples each pyramid |
| 154 | layer by half (if you want to use a finer image pyramid then just change the |
| 155 | template argument to a larger value. For example, using pyramid_down<5> would |
| 156 | downsample each layer by a ratio of 5 to 4). |
| 157 | |
| 158 | Finally, some of the feature extraction zones are allowed to move freely within the |
| 159 | object box. This means that when we are sliding the classifier over an image, some |
nothing calls this directly
no test coverage detected