-
Notifications
You must be signed in to change notification settings - Fork 2
/
Copy pathagglomerative_clustering.cpp
226 lines (197 loc) · 11.3 KB
/
agglomerative_clustering.cpp
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
#include "agglomerative_clustering.h"
void write(FileStorage& fs, const std::string&, const HCluster& x)
{
x.write(fs);
}
void read(const FileNode& node, HCluster& x, const HCluster& default_value)
{
if(node.empty())
x = default_value;
else
x.read(node);
}
std::string xml_path("../trained_boost_groups.xml");
HierarchicalClustering::HierarchicalClustering(vector<Region> &_regions): regions(_regions)
{
#ifndef _TRAIN_
boost = StatModel::load<Boost>(xml_path.c_str());
cerr << "Classifier loaded: " << xml_path << endl;
if( boost.empty() )
{
cerr << "Could not read the classifier: " << xml_path << endl;
CV_Error(Error::StsBadArg, "Could not read the default classifier!");
}
#endif
}
//For feature space
void HierarchicalClustering::operator()(t_float *data, unsigned int num, int dim, unsigned char method, unsigned char metric, vector<HCluster> &merge_info, float x_coord_mult, cv::Size imsize)
{
t_float *Z = (t_float*)malloc(((num-1)*4) * sizeof(t_float)); // we need 4 floats foreach merge
linkage_vector(data, (int)num, dim, Z, method, metric);
build_merge_info(Z, data, (int)num, dim, merge_info, x_coord_mult, imsize);
free(Z);
}
void HierarchicalClustering::build_merge_info(t_float *Z, t_float *X, int N, int dim, vector<HCluster> &merge_info, float x_coord_mult, cv::Size imsize)
{
int this_node = 0;
// walk the whole dendogram
for (int i=0; i<(N-1)*4; i=i+4)
{
HCluster cluster;
cluster.childs.push_back(this_node);
cluster.num_elem = Z[i+3]; //number of elements
int node1 = Z[i];
int node2 = Z[i+1];
float dist = Z[i+2];
if (node1<N) // child node1 is a single region
{
cluster.elements.push_back((int)node1);
cluster.elements_intensities.push_back(regions[node1].intensity_mean_);
cluster.elements_b_intensities.push_back(regions[node1].boundary_intensity_mean_);
cluster.elements_strokes.push_back(regions[node1].stroke_mean_);
cluster.elements_gradients.push_back(regions[node1].gradient_mean_);
cluster.elements_diameters.push_back(max(regions[node1].bbox_.width,regions[node1].bbox_.height));
cluster.rect = regions.at(node1).bbox_;
//vector<float> point;
//for (int n=0; n<dim; n++)
// point.push_back(X[node1*dim+n]);
cluster.points.push_back(Point((int)((X[node1*dim]/x_coord_mult)* imsize.width),(int)(X[node1*dim+1]*imsize.height)));
}
else // child node1 is a cluster
{
cluster.childs.insert(cluster.childs.end(),
merge_info.at(node1-N).childs.begin(),
merge_info.at(node1-N).childs.end());
cluster.points.insert(cluster.points.end(),
merge_info.at(node1-N).points.begin(),
merge_info.at(node1-N).points.end());
cluster.elements.insert(cluster.elements.end(),
merge_info.at(node1-N).elements.begin(),
merge_info.at(node1-N).elements.end());
cluster.elements_intensities.insert(cluster.elements_intensities.end(),
merge_info.at(node1-N).elements_intensities.begin(),
merge_info.at(node1-N).elements_intensities.end());
cluster.elements_b_intensities.insert(cluster.elements_b_intensities.end(),
merge_info.at(node1-N).elements_b_intensities.begin(),
merge_info.at(node1-N).elements_b_intensities.end());
cluster.elements_strokes.insert(cluster.elements_strokes.end(),
merge_info.at(node1-N).elements_strokes.begin(),
merge_info.at(node1-N).elements_strokes.end());
cluster.elements_gradients.insert(cluster.elements_gradients.end(),
merge_info.at(node1-N).elements_gradients.begin(),
merge_info.at(node1-N).elements_gradients.end());
cluster.elements_diameters.insert(cluster.elements_diameters.end(),
merge_info.at(node1-N).elements_diameters.begin(),
merge_info.at(node1-N).elements_diameters.end());
cluster.rect = merge_info.at(node1-N).rect;
}
if (node2<N) // child node2 is a single region
{
//vector<float> point;
//for (int n=0; n<dim; n++)
// point.push_back(X[node2*dim+n]);
//cluster.points.push_back(point);
cluster.points.push_back(Point((int)((X[node2*dim]/x_coord_mult)* imsize.width),(int)(X[node2*dim+1]*imsize.height)));
cluster.elements.push_back((int)node2);
cluster.elements_intensities.push_back(regions[node2].intensity_mean_);
cluster.elements_b_intensities.push_back(regions[node2].boundary_intensity_mean_);
cluster.elements_strokes.push_back(regions[node2].stroke_mean_);
cluster.elements_gradients.push_back(regions[node2].gradient_mean_);
cluster.elements_diameters.push_back(max(regions[node2].bbox_.width,regions[node2].bbox_.height));
cluster.rect = cluster.rect | regions.at(node2).bbox_; // min. area rect containing node 1 and node2
}
else // child node2 is a cluster
{
cluster.childs.insert(cluster.childs.end(),
merge_info.at(node2-N).childs.begin(),
merge_info.at(node2-N).childs.end());
cluster.points.insert(cluster.points.end(),
merge_info.at(node2-N).points.begin(),
merge_info.at(node2-N).points.end());
cluster.elements.insert(cluster.elements.end(),
merge_info.at(node2-N).elements.begin(),
merge_info.at(node2-N).elements.end());
cluster.elements_intensities.insert(cluster.elements_intensities.end(),
merge_info.at(node2-N).elements_intensities.begin(),
merge_info.at(node2-N).elements_intensities.end());
cluster.elements_b_intensities.insert(cluster.elements_b_intensities.end(),
merge_info.at(node2-N).elements_b_intensities.begin(),
merge_info.at(node2-N).elements_b_intensities.end());
cluster.elements_strokes.insert(cluster.elements_strokes.end(),
merge_info.at(node2-N).elements_strokes.begin(),
merge_info.at(node2-N).elements_strokes.end());
cluster.elements_gradients.insert(cluster.elements_gradients.end(),
merge_info.at(node2-N).elements_gradients.begin(),
merge_info.at(node2-N).elements_gradients.end());
cluster.elements_diameters.insert(cluster.elements_diameters.end(),
merge_info.at(node2-N).elements_diameters.begin(),
merge_info.at(node2-N).elements_diameters.end());
cluster.rect = cluster.rect | merge_info.at(node2-N).rect; // min. area rect containing node 1 and node2
}
if ( ((node1<N) && (cluster.rect == regions.at(node1).bbox_)) ||
((node1>N) && (cluster.rect == merge_info.at(node1-N).rect)) )
cluster.inherit_cnn_probability = 1;
else if ( ((node2<N) && (cluster.rect == regions.at(node2).bbox_)) ||
((node2>N) && (cluster.rect == merge_info.at(node2-N).rect)) )
cluster.inherit_cnn_probability = 2;
else
cluster.inherit_cnn_probability = 0;
cluster.node1 = node1-N;
cluster.node2 = node2-N;
//Minibox mb;
//for (int i=0; i<cluster.points.size(); i++)
// mb.check_in(&cluster.points.at(i));
//long double volume = mb.volume();
Rect centers_rect = boundingRect(cluster.points);
long double volume = (long double)centers_rect.area()/(imsize.area());
long double ext_volume = (long double)cluster.rect.area()/(imsize.area());
if (volume >= 1) volume = 0.999999; //TODO this may never happen!!!
if (volume == 0) volume = 0.000001; //TODO is this the minimum we can get? // better if we just quantize to a given grid of possible volumes ...
cluster.nfa = (int)NFA( N, cluster.points.size(), (double) volume, 0); //this uses an approximation for the nfa calculations (faster)
/* predict group class with boost */
int nfa2 = (int)NFA( N, cluster.points.size(), (double) ext_volume, 0); //this uses an approximation for the nfa calculations (faster)
cluster.feature_vector.push_back(cluster.elements.size());
cluster.feature_vector.push_back(cluster.nfa);
cluster.feature_vector.push_back(nfa2);
if (cluster.nfa == 0)
cluster.feature_vector.push_back(-1.0);
else
cluster.feature_vector.push_back((float)nfa2/cluster.nfa);
cluster.feature_vector.push_back((float)(volume/ext_volume));
cluster.feature_vector.push_back((float)centers_rect.width/cluster.rect.width);
int left_diff = centers_rect.x - cluster.rect.x;
int right_diff = (cluster.rect.x+cluster.rect.width) - (centers_rect.x+centers_rect.width);
cluster.feature_vector.push_back((float)left_diff/cluster.rect.width);
cluster.feature_vector.push_back((float)right_diff/cluster.rect.width);
if (max(left_diff,right_diff) == 0)
cluster.feature_vector.push_back(1.0);
else
cluster.feature_vector.push_back((float)min(left_diff,right_diff)/max(left_diff,right_diff));
cluster.feature_vector.push_back((float)centers_rect.height/cluster.rect.height);
int top_diff = centers_rect.y - cluster.rect.y;
int bottom_diff = (cluster.rect.y+cluster.rect.height) - (centers_rect.y+centers_rect.height);
cluster.feature_vector.push_back((float)top_diff/cluster.rect.height);
cluster.feature_vector.push_back((float)bottom_diff/cluster.rect.height);
if (max(top_diff,bottom_diff) == 0)
cluster.feature_vector.push_back(1.0);
else
cluster.feature_vector.push_back((float)min(top_diff,bottom_diff)/max(top_diff,bottom_diff));
Scalar mean,std;
meanStdDev( cluster.elements_diameters, mean, std );
cluster.feature_vector.push_back(std[0]/mean[0]);
meanStdDev( cluster.elements_strokes, mean, std );
cluster.feature_vector.push_back(std[0]/mean[0]);
meanStdDev( cluster.elements_gradients, mean, std );
cluster.feature_vector.push_back(std[0]);
meanStdDev( cluster.elements_intensities, mean, std );
cluster.feature_vector.push_back(std[0]);
meanStdDev( cluster.elements_b_intensities, mean, std );
cluster.feature_vector.push_back(std[0]);
#ifndef _TRAIN_
float votes_group = boost->predict( Mat(cluster.feature_vector), noArray(), DTrees::PREDICT_SUM | StatModel::RAW_OUTPUT);
cluster.probability = (double)1-(double)1/(1+exp(-2*votes_group));
#endif
merge_info.push_back(cluster);
this_node++;
}
}