Peter M. Groen / openbr

Commit ad88ddc7992e5ad6a48dba0d1362667318e8058e

Authored by Charles Otto
1 parent 93733b79

Faster MatchProbability training 

Parallelize KDE lookup table building, also introduce a flag, and
don't do KDE training at all, if Gaussians will instead be used to
model the score distributions
Inline Side-by-side
Showing 1 changed file with 29 additions and 8 deletions
openbr/plugins/quality.cpp
  View file @ad88ddc
... ... @@ -77,6 +77,12 @@ class ImpostorUniquenessMeasureTransform : public Transform
77 77  
78 78 BR_REGISTER(Transform, ImpostorUniquenessMeasureTransform)
79 79  
  80 +
  81 +float KDEPointer(const QList<float> *scores, double x, double h)
  82 +{
  83 + return Common::KernelDensityEstimation(*scores, x, h);
  84 +}
  85 +
80 86 /* Kernel Density Estimator */
81 87 struct KDE
82 88 {
... ... @@ -85,20 +91,35 @@ struct KDE
85 91 QList<float> bins;
86 92  
87 93 KDE() : min(0), max(1), mean(0), stddev(1) {}
88   - KDE(const QList<float> &scores)
  94 +
  95 + KDE(const QList<float> &scores, bool trainKDE)
89 96 {
90 97 Common::MinMax(scores, &min, &max);
91 98 Common::MeanStdDev(scores, &mean, &stddev);
  99 +
  100 + if (!trainKDE)
  101 + return;
  102 +
92 103 double h = Common::KernelDensityBandwidth(scores);
93 104 const int size = 255;
94 105 bins.reserve(size);
95   - for (int i=0; i<size; i++)
96   - bins.append(Common::KernelDensityEstimation(scores, min + (max-min)*i/(size-1), h));
  106 +
  107 + QFutureSynchronizer<float> futures;
  108 +
  109 + for (int i=0; i < size; i++)
  110 + futures.addFuture(QtConcurrent::run(KDEPointer, &scores, min + (max-min)*i/(size-1), h));
  111 + futures.waitForFinished();
  112 +
  113 + foreach(const QFuture<float> & future, futures.futures())
  114 + bins.append(future.result());
97 115 }
98 116  
99 117 float operator()(float score, bool gaussian = true) const
100 118 {
101 119 if (gaussian) return 1/(stddev*sqrt(2*CV_PI))*exp(-0.5*pow((score-mean)/stddev, 2));
  120 + if (bins.empty())
  121 + return -std::numeric_limits<float>::max();
  122 +
102 123 if (score <= min) return bins.first();
103 124 if (score >= max) return bins.last();
104 125 const float x = (score-min)/(max-min)*bins.size();
... ... @@ -123,8 +144,8 @@ struct MP
123 144 {
124 145 KDE genuine, impostor;
125 146 MP() {}
126   - MP(const QList<float> &genuineScores, const QList<float> &impostorScores)
127   - : genuine(genuineScores), impostor(impostorScores) {}
  147 + MP(const QList<float> &genuineScores, const QList<float> &impostorScores, bool trainKDE)
  148 + : genuine(genuineScores, trainKDE), impostor(impostorScores, trainKDE) {}
128 149 float operator()(float score, bool gaussian = true) const
129 150 {
130 151 const float g = genuine(score, gaussian);
... ... @@ -165,7 +186,7 @@ class MatchProbabilityDistance : public Distance
165 186 const QList<int> labels = src.indexProperty(inputVariable);
166 187 QScopedPointer<MatrixOutput> matrixOutput(MatrixOutput::make(FileList(src.size()), FileList(src.size())));
167 188 distance->compare(src, src, matrixOutput.data());
168   -
  189 +
169 190 QList<float> genuineScores, impostorScores;
170 191 genuineScores.reserve(labels.size());
171 192 impostorScores.reserve(labels.size()*labels.size());
... ... @@ -178,8 +199,8 @@ class MatchProbabilityDistance : public Distance
178 199 else impostorScores.append(score);
179 200 }
180 201 }
181   -
182   - mp = MP(genuineScores, impostorScores);
  202 +
  203 + mp = MP(genuineScores, impostorScores, !gaussian);
183 204 }
184 205  
185 206 float compare(const Template &target, const Template &query) const
... ...