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
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
// This file is part of OpenCV project.
// It is subject to the license terms in the LICENSE file found in the top-level directory
// of this distribution and at http://opencv.org/license.html.
#include "precomp.hpp"
#include "face_alignmentimpl.hpp"
using namespace std;
namespace cv{
namespace face{
//Threading helper classes
class doSum : public ParallelLoopBody
{
public:
doSum(vector<training_sample>* samples_,vector<Point2f>* sum_) :
samples(samples_),
sum(sum_)
{
}
virtual void operator()( const Range& range) const CV_OVERRIDE
{
for (int j = range.start; j < range.end; ++j){
for(unsigned long k=0;k<(*samples)[j].shapeResiduals.size();k++){
(*sum)[k]=(*sum)[k]+(*samples)[j].shapeResiduals[k];
}
}
}
private:
vector<training_sample>* samples;
vector<Point2f>* sum;
};
class modifySamples : public ParallelLoopBody
{
public:
modifySamples(vector<training_sample>* samples_,vector<Point2f>* temp_) :
samples(samples_),
temp(temp_)
{
}
virtual void operator()( const Range& range) const CV_OVERRIDE
{
for (int j = range.start; j < range.end; ++j){
for(unsigned long k=0;k<(*samples)[j].shapeResiduals.size();k++){
(*samples)[j].shapeResiduals[k]=(*samples)[j].shapeResiduals[k]-(*temp)[k];
(*samples)[j].current_shape[k]=(*samples)[j].actual_shape[k]-(*samples)[j].shapeResiduals[k];
}
}
}
private:
vector<training_sample>* samples;
vector<Point2f>* temp;
};
class splitSamples : public ParallelLoopBody
{
public:
splitSamples(vector<training_sample>* samples_,vector< vector<Point2f> >* leftsumresiduals_,vector<unsigned long>* left_count_,unsigned long* num_test_splits_,vector<splitr>* feats_) :
samples(samples_),
leftsumresiduals(leftsumresiduals_),
left_count(left_count_),
num_test_splits(num_test_splits_),
feats(feats_)
{
}
virtual void operator()( const Range& range) const CV_OVERRIDE
{
for (int i = range.start; i < range.end; ++i){
for(unsigned long j=0;j<*(num_test_splits);j++){
(*left_count)[j]++;
if ((float)(*samples)[i].pixel_intensities[(unsigned long)(*feats)[j].index1] - (float)(*samples)[i].pixel_intensities[(unsigned long)(*feats)[j].index2] > (*feats)[j].thresh){
for(unsigned long k=0;k<(*samples)[i].shapeResiduals.size();k++){
(*leftsumresiduals)[j][k]=(*leftsumresiduals)[j][k]+(*samples)[i].shapeResiduals[k];
}
}
}
}
}
private:
vector<training_sample>* samples;
vector< vector<Point2f> >* leftsumresiduals;
vector<unsigned long>* left_count;
unsigned long* num_test_splits;
vector<splitr>* feats;
};
splitr FacemarkKazemiImpl::getTestSplits(vector<Point2f> pixel_coordinates,int seed)
{
splitr feat;
//generates splits whose probability is above a particular threshold.
//P(u,v)=e^(-distance/lambda) as described in the research paper
//cited above. This helps to select closer pixels hence make efficient
//splits.
double probability;
double check;
RNG rng(seed);
do
{
//select random pixel coordinate
feat.index1 = rng.uniform(0,params.num_test_coordinates);
//select another random coordinate
feat.index2 = rng.uniform(0,params.num_test_coordinates);
Point2f pt = pixel_coordinates[(unsigned long)feat.index1]-pixel_coordinates[(unsigned long)feat.index2];
double distance = sqrt((pt.x*pt.x)+(pt.y*pt.y));
//calculate the probability
probability = exp(-distance/params.lambda);
check = rng.uniform(double(0),double(1));
}
while(check>probability||feat.index1==feat.index2);
feat.thresh =(float)(((rng.uniform(double(0),double(1)))*256 - 128)/2.0);
return feat;
}
bool FacemarkKazemiImpl:: getBestSplit(vector<Point2f> pixel_coordinates, vector<training_sample>& samples,unsigned long start ,
unsigned long end,splitr& split,vector< vector<Point2f> >& sum,long node_no)
{
if(samples[0].shapeResiduals.size()!=samples[0].current_shape.size()){
String error_message = "Error while generating split.Residuals are not complete.Aborting....";
CV_Error(Error::StsBadArg, error_message);
}
//This vector stores the matrices where each matrix represents
//sum of the residuals of shapes of samples which go to the left
//child after split
vector< vector<Point2f> > leftsumresiduals;
leftsumresiduals.resize(params.num_test_splits);
vector<splitr> feats;
//generate random splits and selects the best split amongst them.
for (unsigned long i = 0; i < params.num_test_splits; ++i){
feats.push_back(getTestSplits(pixel_coordinates,i+(int)time(0)));
leftsumresiduals[i].resize(samples[0].shapeResiduals.size());
}
vector<unsigned long> left_count;
left_count.resize(params.num_test_splits);
parallel_for_(Range(start,end),splitSamples(&samples,&leftsumresiduals,&left_count,¶ms.num_test_splits,&feats));
//Selecting the best split
double best_score =-1;
unsigned long best_feat = 0;
double score = -1;
vector<Point2f> right_sum;
right_sum.resize(sum[node_no].size());
vector<Point2f> left_sum;
left_sum.resize(sum[node_no].size());
unsigned long right_cnt;
for(unsigned long i=0;i<leftsumresiduals.size();i++){
right_cnt = (end-start+1)-left_count[i];
for(unsigned long k=0;k<leftsumresiduals[i].size();k++){
if (right_cnt!=0){
right_sum[k].x=(sum[node_no][k].x-leftsumresiduals[i][k].x)/right_cnt;
right_sum[k].y=(sum[node_no][k].y-leftsumresiduals[i][k].y)/right_cnt;
}
else
right_sum[k]=Point2f(0,0);
if(left_count[i]!=0){
left_sum[k].x=leftsumresiduals[i][k].x/left_count[i];
left_sum[k].y=leftsumresiduals[i][k].y/left_count[i];
}
else
left_sum[k]=Point2f(0,0);
}
Point2f pt1(0,0);
Point2f pt2(0,0);
for(unsigned long k=0;k<left_sum.size();k++){
pt1.x = pt1.x + (float)(left_sum[k].x*left_sum[k].x);
pt2.x = pt2.x + (float)(right_sum[k].x*right_sum[k].x);
pt1.y = pt1.y + (float)(left_sum[k].y*left_sum[k].y);
pt2.y = pt2.y + (float)(right_sum[k].y*right_sum[k].y);
}
score = (double)sqrt(pt1.x+pt1.y)*(double)left_count[i] + (double)sqrt(pt2.x+pt2.y)*(double)right_cnt;
if(score > best_score){
best_score = score;
best_feat = i;
}
}
sum[2*node_no+1] = leftsumresiduals[best_feat];
sum[2*node_no+2].resize(sum[node_no].size());
for(unsigned long k=0;k<sum[node_no].size();k++){
sum[2*node_no+2][k].x = sum[node_no][k].x-sum[2*node_no+1][k].x;
sum[2*node_no+2][k].y = sum[node_no][k].y-sum[2*node_no+1][k].y;
}
split = feats[best_feat];
return true;
}
void FacemarkKazemiImpl::createSplitNode(regtree& tree, splitr split,long node_no){
tree_node node;
node.split = split;
node.leaf.clear();
tree.nodes[node_no]=node;
}
void FacemarkKazemiImpl::createLeafNode(regtree& tree,long node_no,vector<Point2f> assign){
tree_node node;
node.split.index1 = (uint64_t)(-1);
node.split.index2 = (uint64_t)(-1);
node.leaf = assign;
tree.nodes[node_no] = node;
}
bool FacemarkKazemiImpl :: generateSplit(queue<node_info>& curr,vector<Point2f> pixel_coordinates, vector<training_sample>& samples,
splitr &split , vector< vector<Point2f> >& sum){
long start = curr.front().index1;
long end = curr.front().index2;
long _depth = curr.front().depth;
long node_no =curr.front().node_no;
curr.pop();
if(start == end)
return false;
getBestSplit(pixel_coordinates,samples,start,end,split,sum,node_no);
long mid = divideSamples(split, samples, start, end);
//cout<<mid<<endl;
if(mid==start||mid==end+1)
return false;
node_info _left,_right;
_left.index1 = start;
_left.index2 = mid-1;
_left.depth = _depth +1;
_left.node_no = 2*node_no+1;
_right.index1 = mid;
_right.index2 = end;
_right.depth = _depth +1;
_right.node_no = 2*node_no+2;
curr.push(_left);
curr.push(_right);
return true;
}
bool FacemarkKazemiImpl :: buildRegtree(regtree& tree,vector<training_sample>& samples,vector<Point2f> pixel_coordinates){
if(samples.size()==0){
String error_message = "Error while building regression tree.Empty samples. Aborting....";
CV_Error(Error::StsBadArg, error_message);
}
if(pixel_coordinates.size()==0){
String error_message = "Error while building regression tree.No pixel coordinates. Aborting....";
CV_Error(Error::StsBadArg, error_message);
}
queue<node_info> curr;
node_info parent;
vector< vector<Point2f> > sum;
const long numNodes =(long)pow(2,params.tree_depth);
const long numSplitNodes = numNodes/2 - 1;
sum.resize(numNodes+1);
sum[0].resize(samples[0].shapeResiduals.size());
parallel_for_(cv::Range(0,(int)samples.size()), doSum(&(samples),&(sum[0])));
parent.index1=0;
parent.index2=(long)samples.size()-1;
parent.node_no=0;
parent.depth=0;
curr.push(parent);
tree.nodes.resize(numNodes+1);
//Total number of split nodes
while(!curr.empty()){
pair<long,long> range= make_pair(curr.front().index1,curr.front().index2);
long node_no = curr.front().node_no;
splitr split;
//generate a split
if(node_no<=numSplitNodes){
if(generateSplit(curr,pixel_coordinates,samples,split,sum)){
createSplitNode(tree,split,node_no);
}
//create leaf
else{
long count = range.second-range.first +1;
vector<Point2f> temp;
temp.resize(samples[range.first].shapeResiduals.size());
parallel_for_(Range(range.first, range.second), doSum(&(samples),&(temp)));
for(unsigned long k=0;k<temp.size();k++){
temp[k].x=(temp[k].x/count)*params.learning_rate;
temp[k].y=(temp[k].y/count)*params.learning_rate;
}
// Modify current shape according to the weak learners.
parallel_for_(Range(range.first,range.second), modifySamples(&(samples),&(temp)));
createLeafNode(tree,node_no,temp);
}
}
else
{
unsigned long count = range.second-range.first +1;
vector<Point2f> temp;
temp.resize(samples[range.first].shapeResiduals.size());
parallel_for_(Range(range.first, range.second), doSum(&(samples),&(temp)));
for(unsigned long k=0;k<temp.size();k++){
temp[k].x=(temp[k].x/count)*params.learning_rate;
temp[k].y=(temp[k].y/count)*params.learning_rate;
}
// Modify current shape according to the weak learners.
parallel_for_(Range(range.first,range.second), modifySamples(&(samples),&(temp)));
createLeafNode(tree,node_no,temp);
curr.pop();
}
}
return true;
}
unsigned long FacemarkKazemiImpl::divideSamples (splitr split,vector<training_sample>& samples,unsigned long start,unsigned long end)
{
if(samples.size()==0){
String error_message = "Error while dividing samples. Sample array empty. Aborting....";
CV_Error(Error::StsBadArg, error_message);
}
unsigned long i = start;
training_sample temp;
//partition samples according to the split
for (unsigned long j = start; j < end; ++j)
{
if ((float)samples[j].pixel_intensities[(unsigned long)split.index1] - (float)samples[j].pixel_intensities[(unsigned long)split.index2] > split.thresh)
{
temp=samples[i];
samples[i]=samples[j];
samples[j]=temp;
++i;
}
}
return i;
}
}//cv
}//face